2

I found that in Bayesian Lasso commonly $\beta \sim N(0,\sigma^2*diag(\tau))$ and $\sigma,\tau \sim \pi(\sigma,\tau)$ is used. Whereas $\pi(\cdot)$ is a product of Laplace distributions.

Is it possible to calculate the conditional density $\tau|\sigma,\beta$? Or do I have to run another Metropolis-Hasting within the sampler in that case?

I have a good draw for $\beta$ and now I want to update $\tau$ accordingly. Here the common Lasso setting does not work.

Xi'an
  • 90,397
  • 9
  • 157
  • 575
JohnScott
  • 125
  • 6
  • What is the point of having both $\sigma$ and $\tau$ as hyperparameters? – Xi'an Nov 06 '15 at 17:39
  • I am not able to run a full gibbs sampling. At the point I have a new $\beta$ i need to update $\tau$ and so on... – JohnScott Nov 06 '15 at 17:53
  • My point is rather that your prior is $\beta \sim N(0,\sigma^2\tau I)$ so I cannot distinguish $\tau$ from $\sigma$ in the model, all that matters is $\tau\sigma^2$. – Xi'an Nov 06 '15 at 17:56
  • I see, you are right. But also for $\tau\sigma^2$ I need the conditional. By the way: For instance in the Bayesian Elastic Net there are at least four parameters to set $\rho_{1/2}$ and $r_{1/2}$. Do you know a reference for that? Sorry, for all this odd questions... – JohnScott Nov 06 '15 at 18:49
  • I think [this thread](http://stats.stackexchange.com/q/175458/7224) replies to your question. – Xi'an Nov 06 '15 at 19:44

0 Answers0