4
$\begingroup$

"The canonical link of Gamma GLM is $g(x)=1/x$ is often not very practical. Log-link is more appropriated in most cases." One reason I can think of is that log-link makes sure $\mu$, the mean, is always greater than zero. (please correct me if I'm wrong.) But I believe there are other considerations.

So why people often choose the log-link over the canonical one? Is it related to the fact that $\log$ transformation stabilizes data variance?

$\endgroup$
3
  • $\begingroup$ who are you quoting? $\endgroup$
    – Glen_b
    Aug 25, 2020 at 6:07
  • 1
    $\begingroup$ Quoting my class instructor, Prof. Hans Muller. $\endgroup$
    – WCMC
    Aug 25, 2020 at 6:23
  • $\begingroup$ The prime reason would be that you expect that the logs of the conditional means are linear in the supplied predictors. $\endgroup$
    – Glen_b
    Aug 25, 2020 at 6:59

1 Answer 1

5
$\begingroup$

A good question. I see the following reasons:

  • A log link produces a multiplicative model on the original scale and is thus easy to interprete. This is not true for the canonical link of the Gamma GLM.

  • In insurance pricing, the expected claim amount $E(L)$ is often decomposed into $E(L) = E(F)E(S)$, where $F$ is the claim frequency and $S$ the claim height. If both $E(F)$ and $E(S)$ are modeled by GLMs with log link, the coefficients of both models can be easily combined to determine the effects on the expected loss $E(L)$.

  • Not using the canonical link introduces a prediction bias on the original scale, i.e. the average prediction differs from the average response. This would be a reason against the log link. However, in my experience, this bias is small with the log link and can easily be fixed by a multiplicative correction factor.

$\endgroup$
1
  • $\begingroup$ Thanks. Can you please give more detail on "decomposed into a product of a claim count regression and a claim height regression"? And what is "prediction bias"? $\endgroup$
    – WCMC
    Aug 24, 2020 at 16:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.