Consider $Y_1, \dots, Y_n$ independent from the Gamma distribution. For $y > 0$: $$\begin{align} f(y \mid \alpha, \beta) &= \dfrac{1}{\beta^{\alpha}\Gamma(\alpha)}y^{\alpha-1}e^{-y/\beta} \\ &= \exp\left[-\log(\beta^{\alpha}\Gamma(\alpha))+(\alpha-1)\log(y)-\dfrac{y}{\beta} \right] \\ &= \exp\left[y\left(\dfrac{-1}{\beta} \right)-\log(\beta^{\alpha}\Gamma(\alpha))+(\alpha-1)\log(y) \right] \end{align}$$ I define an exponential dispersion family as any distribution whose PMF/PDF is $$f(y \mid \boldsymbol\theta) = \exp\left\{\phi[y\theta - b(\theta)] + c(y, \phi) \right\}\text{, } y \in \Omega$$ where $\Omega$ is in the support of a random variable $Y$ in the family.

Method 1:

Therefore, $Y_i$ is of the exponential dispersion family with (assuming $\alpha_i$ is known) \begin{align} \phi &= 1 \\ \theta_i &= \dfrac{-1}{\beta_i} \\ b(\theta_i) &= \log[\beta_i^{\alpha_i}\Gamma(\alpha_i)] \\ &= \log[(-\theta_i)^{-\alpha_i}\Gamma(\alpha_i)] \\ &= -\alpha_i\log(-\theta_i) + \log \Gamma(\alpha_i) \\ c(\phi, y_i) &= (\alpha_i-1)\log(y_i)\text{.} \end{align} It follows that \begin{equation} \mu_i = b^{\prime}(\theta_i) = \dfrac{-\alpha_i}{-\theta_i}(-1) = \dfrac{-\alpha_i}{\theta_i}\text{.} \end{equation} and our canonical link is

$$g(\mu_i) = \theta_i = -\dfrac{\alpha_i}{\mu_i}\tag{*}$$

Method 2:

Credit due to http://civil.colorado.edu/~balajir/CVEN6833/lectures/GammaGLM-01.pdf.

Let $\theta = -\dfrac{1}{\alpha\beta}$. Then rewrite $$\begin{align}f(y \mid \alpha, \beta) &= \exp\left[y\left(\dfrac{-1}{\beta} \right)-\log(\beta^{\alpha}\Gamma(\alpha))+(\alpha-1)\log(y) \right] \\ &= \exp\left[y\left(\dfrac{-1}{\beta} \right)-\alpha\log(\beta)-\log\Gamma(\alpha)+(\alpha-1)\log(y) \right] \\ &= \exp\left[y\alpha\theta-\alpha\log(\theta^{-1})-\log\Gamma(\alpha)+(\alpha-1)\log(y) \right] \\ &= \exp\left\{\alpha[y\theta-\log(\theta^{-1})]-\log\Gamma(\alpha)+(\alpha-1)\log(y) \right\} \end{align}$$ in which case, since $\phi$ cannot vary depending on $i$, \begin{align} \phi &= \alpha \\ \theta_i &= -\dfrac{1}{\alpha\beta_i} \\ b(\theta_i) &= \log(\theta_i^{-1}) = -\log(\theta_i)\\ c(\phi, y_i) &= (\alpha-1)\log(y_i)-\log\Gamma(\alpha)\text{.} \end{align} In this case, $b^{\prime}(\theta_i) =-\dfrac{1}{\theta_i} = \mu_i$, which means that $$g(\mu_i) = \theta_i = -\dfrac{1}{\mu_i}\text{.}\tag{**}$$

Wikipedia cites (**) as the canonical link function. Why is (**) preferable over (*) for the canonical link function?

What makes this even more confusing is that http://civil.colorado.edu/~balajir/CVEN6833/lectures/GammaGLM-01.pdf says that the canonical link function is $\dfrac{1}{\mu_i}$ (also mentioned in McCullagh and Nelder).

First main question: Are canonical link functions unique, when provided a random component?

Second main question: Is there only supposed to be one canonical link function for each random component? If so, what is wrong with the work above?


1 Answer 1


The systematic component of the GLM is $$g(\mu_i) = \mathbf{x}_i^{\prime}\boldsymbol\beta\text{.}$$ As long as we do not place any restrictions on the $\boldsymbol\beta$ coefficients, without loss of generality, we may ignore all constants (with respect to $\mu_i$) of proportionality (these constants will just be absorbed into $\boldsymbol\beta$). Thus, any of the link functions above will work as a canonical link function, but $g(\mu_i) = \dfrac{1}{\mu_i}$ is the simplest one to work with.


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.