Suppose that $x$ is an observable scalar variable and $$ x \thicksim N(W\mu_0,W^2\sigma_0^2) $$ Where $W$ is a parameter that must be estimated from data, and $\mu_0$ , $\sigma_0$ are known constants.

I want to solve the problem in Bayesian framework, so I consider $W$ as a random variable, and I have to choose a prior probability distribution for it.

What is a suitable conjugate prior for W in my problem? (I prefer the priors which belong to a known family of distributions).

  • $\begingroup$ Could you give some information about what you are trying to model? $\endgroup$ – Pieter May 18 '16 at 19:34

The natural conjugate prior for this model is $$ p(W) \propto W^{-a}\exp\left(-\frac{b}{W^2}+\frac{c}{W}\right)$$ and its posterior is $$ p(W|x) \propto W^{-(a+1)}\exp\left(-\frac{(b+x^2/[2\sigma_0^2])}{W^2}+\frac{(c+x\mu_0/\sigma_0^2)}{W}\right).$$ It is not from a family of distributions that I know and I have no idea if, or for what parameter settings and data, this will be valid probability density.

This problem is related to the generalized inverse Gaussian distribution which is the conjugate prior for the model $$ x|W \sim N(\alpha+\beta W,W) $$ the difference is that $W$ shows up in both the mean and the variance rather than $W^2$ in the variance.

  • $\begingroup$ Thanks. Actually, I need to draw a sample from this distribution in my C++ program. Can you suggest an efficient method? $\endgroup$ – A.Ahmadian May 20 '16 at 14:12

Your Answer

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

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