In my lecture notes I have that the distribution of a random variable $Y$ is said to be in the exponential family if it can be written as $f(y;\theta)=exp(a(y)b(\theta)+c(\theta)+d(y))$, where $a,b,c$ and $d$ are fixed functions. If we have (in the exponential family form) that $a(y)=y$, then the distribution is said to be in canonical form and then $b(\theta)$ is the called the natural parameter and $b$ is called the natural link function. Then we have shown that for a random variable $Y$ in exponential family form we have $\mathbb{E}(a(Y))=\frac{-c'(\theta)}{b'(\theta)}$ and $var(a(Y))=\frac{b''(\theta)c'(\theta)-c''(\theta)b'(\theta)}{[b'(\theta)]^{3}}$.

I am now working on the following problem: the gamma pdf of r.v. $Y$ is given by

$\begin{equation} f(y) = (s^{a}\Gamma (a))^{-1}y^{a-1}e^{y/s} \end{equation}$,

where $y \geq 0$, $s$ is the scale parameter, $a$ the shape parameter. The first question asks me to reparameterise this pdf by setting $a=1/\phi$ and $s=\mu\phi$, and hence show that it is a member of the exponential family. So after introducing $a=1/\phi$ and $s=\mu\phi$ into the pdf and rearranging I get

$ f(y) = exp((\frac{1}{\phi}-1)log(y)-\frac{y}{\mu\phi}-\frac{1}{\phi}log(\mu\phi)-log(\Gamma(1/\phi))=exp(a(y)b(\theta)+c(\theta)+d(y))$,

where $a(y)=y$,$b(\mu)=-\frac{1}{\mu\phi}$,$c(\mu)=-\frac{1}{\phi}log(\mu\phi)$ and $d(y)=(\frac{1}{\phi}-1)log(y) - log(\Gamma(\frac{1}{\phi}))$, where we treat the (dispersion) parameter $\phi$ as a nuisance parameter. Is that correct?

The next question says: deduce that the canonical link for the gamma is $\theta=\frac{1}{\mu}=\eta=\textbf{X}\beta$. So I'm thinking since $f$ is in a canonical form, the canonical parameter is $b(\mu)=-\frac{1}{\mu\phi}$ and so ignoring all the constants of proportionality we have that the canonical link, in its simplest form, is $\frac{1}{\mu}$, as required. Does that make sense? I still don't have a good enough understanding of link/canonical link functions, I'm afraid.

Then the next question asks me to deduce further that the variance function is $b''(\theta)=-1/\theta^{2}=-\mu^{2}$. I don't really know how to do this. Why is this the variance function? What is $\theta$ here? A canonical link? That clearly doesn't make sense. Or is just the dummy variable for our parameter of interest ( $\mu$ in our case )? I've tried to differentiate $b(\mu)$ w.r.t. to $\mu$ but it doesn't work. I'd really appreciate some help.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.