Common Gamma Link Interpretation Errors

Why Link Interpretation Matters

A coefficient in a Gamma generalized linear model describes a change on the model’s link scale. It does not automatically describe an equal-sized change in the response mean. Confusing the two scales can reverse an effect, misstate its size, or conceal a prediction that is not a possible Gamma mean.

Write $\mu_i=E[Y_i\mid x_i]$ for the conditional mean of a Gamma response and $\eta_i=x_i^{\mathsf T}\beta$ for the linear predictor. A Gamma response has positive support and requires $\mu_i>0$. The link specifies the relation between $\eta_i$ and $\mu_i$. This tutorial focuses on interpretation errors involving the log link, the reciprocal-mean link, and fitted-mean validity.

Definition: In a Gamma regression, $\mu_i$ is the conditional response mean, and the model sets $g(\mu_i)=\eta_i$. Under the log link, $g(\mu)=\log\mu$, so $\mu_i=\exp(\eta_i)$. Under the reciprocal-mean link used here, $g(\mu)=1/\mu$, so $\mu_i=1/\eta_i$, which requires $\eta_i>0$. The canonical negative-inverse link instead uses $g(\mu)=-1/\mu$; do not silently switch between these sign conventions.

Log-Link Coefficients Are Not Additive Mean Changes

With the log link, a coefficient changes the logarithm of the mean. Exponentiating the predictor gives the mean itself. Consequently, for a fixed change in a covariate, the model implies a multiplicative change in the mean, not a fixed additive change.

Theorem: Suppose $\log\mu(x)=\eta(x)$, and changing a covariate by $d$ changes the linear predictor by $\beta_jd$, with all other terms held fixed. Then $\mu(x+d)/\mu(x)=\exp(\beta_jd)$, and $\mu(x+d)-\mu(x)=\mu(x)\{\exp(\beta_jd)-1\}$.

Proof. The link equation gives $\mu(x)=\exp\{\eta(x)\}$ and $\mu(x+d)=\exp\{\eta(x)+\beta_jd\}$. Dividing the second expression by the first gives $\exp(\beta_jd)$. Subtracting the original mean from the changed mean gives $\mu(x)\{\exp(\beta_jd)-1\}$. The difference therefore depends on the starting mean, even when the covariate change and coefficient are fixed. $\square$

Worked Example: A Log-Link Coefficient Is a Ratio

Suppose a Gamma model for mean treatment cost has a log link and includes a binary indicator $A$, with coefficient $\beta_A=0.20$. The reference group has fitted mean $\$10{,}000$. Find the fitted mean for $A=1$ and interpret the coefficient.

Changing $A$ from 0 to 1 changes the predictor by $0.20$. The mean ratio is $\exp(0.20)\approx1.2214$. Thus the group with $A=1$ has a fitted mean about 1.2214 times the reference mean, or about 22.14% higher.

The fitted mean is $\$10{,}000\times1.2214\approx\$12{,}214$, a difference of about $\$2{,}214$. The coefficient is not an increase of $\$0.20$, nor is it an increase of $\$2{,}000$. The additive mean difference depends on the reference mean.

For a continuous covariate, the same calculation applies to any specified change $d$, not just one unit. A one-unit increase multiplies the mean by $\exp(\beta_j)$; a three-unit increase multiplies it by $\exp(3\beta_j)$. The derivative $d\mu/dx_j=\beta_j\mu$ describes a local rate of change, not a constant additive change across all starting means. In particular, approximating $\exp(\beta_j)-1$ by $\beta_j$ is only a small-coefficient approximation to the relative change, not an exact mean difference.

Worked Example: A Continuous Covariate and a Specified Change

A Gamma log-link model for a positive biomarker has $\log\mu(x)=\beta_0+0.08x$. At $x=5$, its fitted mean is 10 units. Calculate the fitted mean at $x=8$.

The covariate change is $d=8-5=3$. The predictor increases by $0.08(3)=0.24$, so the mean ratio is $\exp(0.24)\approx1.27125$. The fitted mean at $x=8$ is therefore $10(1.27125)\approx12.7125$ units, an increase of about 2.7125 units.

The coefficient $0.08$ is not an increase of 0.08 biomarker units for each unit of $x$. It gives a one-unit mean ratio of $\exp(0.08)\approx1.0833$; the three-unit change has the larger ratio $\exp(0.24)$.

Reciprocal-Mean Coefficients Act on the Reciprocal Scale

Under the reciprocal-mean link, the linear predictor equals $1/\mu$, not $\mu$. A coefficient is additive in the reciprocal mean. Because taking a reciprocal reverses the order of positive values, an increase in the predictor corresponds to a decrease in the response mean. The size of that decrease depends on the starting predictor and the change.

Theorem: Suppose $1/\mu(x)=\eta(x)>0$, and changing a covariate by $d$ changes the predictor by $\beta_jd$. If both $\eta(x)$ and $\eta(x)+\beta_jd$ are positive, then $\mu(x+d)=1/\{\eta(x)+\beta_jd\}$ and $\mu(x+d)/\mu(x)=\eta(x)/\{\eta(x)+\beta_jd\}$.

Proof. The reciprocal-link definition gives $\mu(x)=1/\eta(x)$. After the covariate change, the predictor is $\eta(x)+\beta_jd$, so the new mean is its reciprocal, $1/\{\eta(x)+\beta_jd\}$. Dividing this by $1/\eta(x)$ gives the stated ratio. Both predictor values must be positive for these reciprocals to be finite positive means. $\square$

Worked Example: Interpreting a Reciprocal-Mean Coefficient

Suppose a model uses $1/\mu=0.25+0.05z$, where the response is a positive duration in days. Find the fitted means at $z=0$ and $z=1$, and interpret the coefficient.

At $z=0$, the predictor is $0.25$, so $\mu=1/0.25=4$ days. At $z=1$, the predictor is $0.25+0.05=0.30$, so $\mu=1/0.30=3.\overline{3}$ days. The mean ratio is $4/(3.\overline{3})=1.2$ in the reverse direction: the $z=0$ mean is 1.2 times the $z=1$ mean. Equivalently, the mean falls by $4-3.\overline{3}=0.\overline{6}$ days, or about 16.7% of its starting value, when $z$ increases from 0 to 1.

The coefficient $0.05$ adds 0.05 to the reciprocal mean. It does not add 0.05 days to the response mean. Here the predictor increases, so its reciprocal—the mean—decreases.

A change in the reciprocal predictor must also preserve its allowed range. In this convention the new predictor must remain strictly positive. A coefficient’s sign alone does not establish whether a proposed prediction is valid: the starting predictor, the covariate value, and the size of the change all matter.

Check Fitted Means Before Interpreting Them

The Gamma response mean must be strictly positive. A model can therefore be inappropriate at a particular covariate pattern if its link and predictor produce a nonpositive or undefined mean. Checking this condition is not a cosmetic step: it determines whether the model’s stated mean is mathematically compatible with a Gamma response.

For a finite real predictor, the log link gives $\mu=\exp(\eta)>0$, so it cannot produce a nonpositive mean. The reciprocal-mean link gives a positive mean only when $\eta>0$; $\eta=0$ has no finite reciprocal, and $\eta<0$ produces a negative value. The identity link gives $\mu=\eta$, so it requires $\eta>0$ at every covariate pattern where a Gamma mean is claimed.

Worked Example: Detecting an Invalid Identity-Link Prediction

An identity-link Gamma model has fitted mean $\mu=2+0.6x$, where $x$ is a centered covariate. Check its fitted means at $x=-4$, $x=0$, and $x=2$.

At $x=-4$, $\mu=2+0.6(-4)=2-2.4=-0.4$. This is invalid for a Gamma response. At $x=0$, $\mu=2$, which is positive and admissible. At $x=2$, $\mu=2+1.2=3.2$, also positive and admissible.

The model’s positive predictions at some values do not repair the invalid prediction at $x=-4$. If that covariate pattern is part of the intended prediction domain, the model cannot provide a valid Gamma mean there. A positive observed response does not make a negative fitted mean valid.

The domain check depends on the link convention. With the reciprocal-mean link $1/\mu=\eta$, validity requires $\eta>0$. With the canonical negative-inverse link $-1/\mu=\eta$, validity instead requires $\eta<0$. In either case, the resulting mean must be finite and strictly positive. State the link equation first, then check the predictor against its domain.

Common Interpretation Errors

Reading a log-link coefficient as an additive mean change. A coefficient is additive in $\log\mu$. Exponentiating the coefficient for the specified covariate change gives a mean ratio; the absolute mean difference also depends on the baseline mean.
Calling the reciprocal predictor the response mean. Under $1/\mu=\eta$, the predictor is $1/\mu$. Recover the response mean by taking its reciprocal, and check that $\eta>0$.
Assuming every link produces an admissible Gamma mean. The identity link can give a nonpositive predictor, and the reciprocal-mean link can give a nonpositive or zero predictor. Check fitted means over the intended covariate domain, including prediction points.
Ignoring the sign convention for inverse links. The equations $1/\mu=\eta$ and $-1/\mu=\eta$ have different predictor domains. Do not carry an inequality or coefficient interpretation from one convention into the other without checking the definition.

Key takeaway: Interpret a coefficient on the scale defined by its link, transform back to the response mean before reporting a mean-scale effect, and verify that every fitted Gamma mean is finite and strictly positive.

Check Your Understanding

Use the stated link equation in each question, and distinguish the predictor scale from the response-mean scale.

Under a log link, a binary covariate has coefficient $-0.30$. What is the fitted mean ratio for covariate value 1 versus 0?
A log-link model has baseline mean 5 and a covariate change that increases the predictor by $0.1$. Why is the mean difference not simply 0.1?
Under $1/\mu=\eta$, the predictor changes from 0.4 to 0.5. Calculate both means and state the direction of change.
For the reciprocal-mean link, what is wrong with predicting from $\eta=0$?
An identity-link Gamma model predicts $\mu=-1$ for one covariate pattern. Does a positive observed response make this prediction admissible? Explain.