Why the Inverse Link Requires a Domain Check
A Gamma regression is used for a positive, continuous response when a Gamma distribution and its mean-variance relationship are suitable. With the inverse link, the model is linear in the reciprocal of the conditional mean, not in the mean itself. This gives coefficients a different interpretation from the multiplicative interpretation in Gamma regression with a log link.
There is an important consequence: a positive Gamma mean requires a positive linear predictor. Unlike a link that maps every real number to a positive mean, the inverse link is defined here only when the predictor is positive. A calculation that produces a zero or negative predictor does not give a valid fitted mean.
Here the Gamma distribution uses the shape-scale parameterization: a Gamma variable with shape \(\alpha\) and scale \(\theta\) has mean \(\alpha\theta\) and variance \(\alpha\theta^2\). Substituting \(\alpha_i=1/\phi\) and \(\theta_i=\phi\mu_i\) gives \(E(Y_i\mid\boldsymbol{x}_i)=(1/\phi)(\phi\mu_i)=\mu_i\) and \(\operatorname{Var}(Y_i\mid\boldsymbol{x}_i)=(1/\phi)(\phi\mu_i)^2=\phi\mu_i^2\). The usual regression setup also treats responses as conditionally independent given their predictors. The link determines the mean; it does not say that an individual response equals its fitted mean.
The domain condition is not optional. If \(\eta_i=0\), its reciprocal is not finite. If \(\eta_i<0\), its reciprocal is negative, which cannot be the mean of a positive Gamma response. Therefore, a model’s predictions should be checked over the covariate values where it will be used, not just at one example row.
How to Interpret Predictor Changes
The linear predictor changes additively with a predictor, but the response mean changes reciprocally. Holding all other predictors fixed, increasing \(x_j\) by \(\Delta\) changes the predictor from \(\eta\) to \(\eta+\beta_j\Delta\). If both are positive, the resulting mean ratio is the old predictor divided by the new predictor.
Thus, a coefficient is not a constant change in the mean, and it is not a constant percentage change in the mean. Its sign does give a useful local direction when the model remains valid: for \(\beta_j>0\), increasing \(x_j\) increases \(\eta\) and decreases \(\mu\); for \(\beta_j<0\), increasing \(x_j\) decreases \(\eta\) and increases \(\mu\), provided the new predictor remains positive. The size of the mean change depends on the current predictor value.
This derivative follows from the chain rule: the derivative of \(1/\eta\) with respect to \(\eta\) is \(-1/\eta^2\), and the derivative of \(\eta\) with respect to \(x_j\), holding other predictors fixed, is \(\beta_j\). The derivative describes a local rate of change, not a finite change over an arbitrary predictor interval. For a finite interval, use the reciprocal formula above.
Worked Example: A Claim Amount and a Risk Score
Method: calculate the linear predictor first, then take its reciprocal to obtain the fitted mean. A hypothetical insurer models claim amount \(Y\), measured in thousands of dollars, using \(\eta=0.05+0.01s\), where \(s\) is a nonnegative risk score. Suppose \(\phi=0.25\). Find the fitted mean at scores 0 and 2, and compare them. The conditional distribution has shape \(1/\phi=4\) and scale \(\phi\mu=0.25\mu\).
At \(s=0\), the predictor is \(\eta=0.05+0.01(0)=0.05\), which is positive. Therefore, \(\mu=1/0.05=20\) thousand dollars. The Gamma scale is \(0.25(20)=5\) thousand dollars, so the conditional response distribution is Gamma with shape 4 and scale 5 thousand dollars.
At \(s=2\), \(\eta=0.05+0.01(2)=0.07\), also positive. The fitted mean is \(\mu=1/0.07=100/7=14.2857\) thousand dollars, and the Gamma scale is \(0.25(100/7)=25/7=3.5714\) thousand dollars. The conditional distribution at this score is Gamma with shape 4 and scale \(25/7\) thousand dollars.
The ratio of the score-2 mean to the score-0 mean is \(0.05/0.07=5/7\), about \(0.7143\). The score-2 fitted mean is therefore about 71.43% of the score-0 mean. Check: the reciprocal of \(14.2857\) is approximately \(0.07\), as required. The mistake avoided is treating the score coefficient \(0.01\) as a direct increase of \(0.01\) thousand dollars in the mean.
Indicator Predictors and Feasible Predictions
For a categorical predictor represented by an indicator, its coefficient changes the reciprocal mean. It does not by itself give the percentage change in the mean. To compare categories, calculate each category’s predictor, verify that both predictors are positive, and then take their reciprocals or their ratio.
Feasibility also depends on the range of the predictors. A model can produce valid fitted means for the observed covariate combinations but an invalid predictor at a new combination. With a linear predictor, checking its minimum over a specified range is often an efficient way to identify this problem. For example, a predictor that decreases as a nonnegative covariate increases may eventually reach zero if extrapolated far enough.
Worked Example: Comparing Two Rating Groups
Method: evaluate both group-specific predictors and compare reciprocal means. A hypothetical model for positive claim-settlement duration \(Y\), measured in days, is \(\eta=0.10+0.02z-0.005x\), where \(z=1\) denotes Group B, \(z=0\) denotes Group A, and \(x\) is experience in years. At \(x=4\), compare the fitted means. Let \(\phi=0.20\), so the Gamma shape is 5 and its scale is \(0.20\mu\).
For Group A, \(z=0\), so \(\eta_A=0.10+0.02(0)-0.005(4)=0.10-0.02=0.08\). This is positive, and the fitted mean is \(\mu_A=1/0.08=12.5\) days. Its Gamma scale is \(0.20(12.5)=2.5\) days.
For Group B, \(z=1\), so \(\eta_B=0.10+0.02(1)-0.005(4)=0.10+0.02-0.02=0.10\). This is positive, and \(\mu_B=1/0.10=10\) days. Its Gamma scale is \(0.20(10)=2\) days. The conditional distributions are therefore Gamma with shape 5 and scales 2.5 days and 2 days for Groups A and B, respectively.
The mean ratio is \(\mu_B/\mu_A=10/12.5=0.8\), so Group B’s fitted mean is 20% lower at this experience level. The positive indicator coefficient raises the predictor from 0.08 to 0.10; it does not raise the mean. Check: both means are positive, and their reciprocals recover the calculated predictors. The common mistake avoided is reading the sign of the coefficient as the direction of the mean effect without accounting for the reciprocal link.
Checking Validity Across a Predictor Range
For a continuous predictor, checking only a single prediction can conceal an invalid part of the model’s intended range. If the model will be used for every combination in a stated range, find the smallest predictor over that range. If the predictor is linear in each covariate, its extrema over a rectangular range occur at the endpoints; indicator variables need to be checked at each category value.
If the minimum predictor is positive, all predictors in that range are positive and their fitted means are finite and positive. If the minimum is zero or negative, at least one requested prediction is invalid. A fitted mean computed at a valid point does not remove this limitation elsewhere.
Worked Example: Verifying a Model Over Its Intended Range
Method: find the minimum and maximum predictor over the specified covariate range before interpreting a prediction. A hypothetical Gamma model for positive equipment-repair time \(Y\), in hours, is \(\eta=0.03+0.004x-0.01z\), where \(0\leq x\leq10\) is a condition score and \(z\in\{0,1\}\) is an equipment class indicator. Let \(\phi=0.25\), giving shape 4 and scale \(0.25\mu\). Determine whether the model is feasible throughout this range and predict the mean at \(x=6,z=1\).
Because the coefficient of \(x\) is positive, the predictor increases as \(x\) increases. Because the coefficient of \(z\) is negative, the predictor is smaller at \(z=1\). The minimum is therefore at \(x=0,z=1\): \(\eta_{\min}=0.03+0.004(0)-0.01(1)=0.02\). The maximum is at \(x=10,z=0\): \(\eta_{\max}=0.03+0.004(10)-0.01(0)=0.07\). Since every predictor in the stated range is at least \(0.02>0\), every fitted mean in that range is finite and positive. The smallest predictor gives the largest mean, \(1/0.02=50\) hours; the largest predictor gives the smallest mean, \(1/0.07=14.2857\) hours.
At \(x=6,z=1\), \(\eta=0.03+0.004(6)-0.01(1)=0.03+0.024-0.01=0.044\). Hence \(\mu=1/0.044=250/11=22.7273\) hours. The Gamma scale is \(0.25(250/11)=62.5/11=5.6818\) hours, so the conditional distribution is Gamma with shape 4 and scale \(62.5/11\) hours. Check: the predicted mean lies between 14.2857 and 50 hours, the endpoint means implied by the predictor range. The mistake avoided is checking just the requested prediction and overlooking invalid predictions elsewhere in the stated range.
Common Mistakes and Practical Interpretation
The inverse link can be applied only after forming the entire linear predictor. Do not take reciprocals of individual coefficients or predictors separately. The model is \(\eta_i=\beta_0+\sum_j\beta_jx_{ij}\), followed by \(\mu_i=1/\eta_i\); it is not a sum of reciprocal terms.
A second frequent error is to interpret \(\eta_i\) as the mean. For example, if \(\eta_i=0.08\), then the fitted mean is \(1/0.08=12.5\), not \(0.08\). The mean and predictor also have reciprocal units: if the response is measured in dollars, the predictor has units of inverse dollars (with the corresponding units for any covariate coefficients).
Finally, a model can be algebraically defined at a row of data but still be unsuitable for a requested prediction if its predictor is nonpositive there. Consider \(\eta=0.06-0.008x\) for a nonnegative covariate \(x\). At \(x=8\), \(\eta=0.06-0.064=-0.004\). Its reciprocal is negative, so this does not yield a valid Gamma mean. This is not a prediction to report as a negative response; it is a domain failure at that covariate value.
Check Your Understanding
For each problem, calculate the predictor before the mean and check its sign. These questions are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.
- A positive response is modeled by \(\eta=0.125+0.025x\). At \(x=3\), calculate the fitted mean and state whether the predictor is valid. (Answer: \(\eta=0.200\), so the fitted mean is \(5\) response units; valid.)
- A binary-group model has \(\eta=0.09-0.01z\), where \(z=0\) is Group A and \(z=1\) is Group B. Calculate both means and the ratio of Group B’s mean to Group A’s mean. (Answer: \(\mu_A=11.1111\), \(\mu_B=12.5\), and \(\mu_B/\mu_A=1.125\); both predictors are positive.)
- For \(\eta=0.04+0.003x\), calculate the fitted mean at \(x=8\). If \(\phi=0.5\), give the shape and scale of the conditional Gamma distribution at that point. (Answer: \(\eta=0.064\), \(\mu=15.625\), shape \(2\), scale \(7.8125\); valid.)
- A model has \(\eta=0.02+0.005x-0.03z\), with \(0\leq x\leq4\) and \(z\in\{0,1\}\). Is it valid over the full range? Identify a covariate combination that shows the issue. (Answer: No; at \(x=0,z=1\), \(\eta=-0.01\), so there is no valid positive fitted mean.)
- A model has \(\eta=0.06-0.008x\), with \(x\geq0\). Find the range of \(x\) for which it gives a finite positive fitted mean. What happens at the boundary? (Answer: \(0\leq x<7.5\); at \(x=7.5\), \(\eta=0\), so the reciprocal is undefined.)