Exercises 8

Exercise 8.1

Zellner’s \(g\) prior for linear regression takes the form \[\begin{align*} \boldsymbol{\beta} \mid \omega, \mathbf{X} &\sim \mathsf{Gauss}_p\left(\boldsymbol{\beta}_0, \frac{n}{\kappa_0}\sigma^2(\mathbf{X}^\top\mathbf{X})^{-1}\right)\\ p(\sigma^2) &\propto \frac{1}{\sigma^2} \end{align*}\] which gives a prior variance proportional to that of the sampling distribution of the ordinary least squares estimator, which is \[\widehat{\boldsymbol{\beta}} \sim \mathsf{Gauss}_p\{\boldsymbol{\beta}, \sigma^2(\mathbf{X}^\top\mathbf{X})^{-1}\}.\] Note that the prior is conditional on the covariates, which are assumed to be exogeneous.

  1. Show that Zellner’s \(g\) prior gives rise to a conjugate prior and derive the posterior mean and covariance.
  2. What is the benefit of this relative to an isotropic prior covariance \(\sigma^2\mathsf{I}_p\) for \(\beta \mid \sigma^2\)?
  3. Give an interpretation to \(\kappa_0\): what happens when \(\kappa_0=1\)? or \(\kappa_0=n\)?