4

Let $X \sim N_q(\mu, \Sigma)$ with a Normal-Inverse-Wishart prior on $(\mu, \Sigma)$ i.e. $(\mu, \Sigma) \sim N_q(\mu | m, \Sigma / k) IW_q(\Sigma | \nu, \Lambda)$ where we are using the Inverse Wishart paramaterization that gives $E [\Sigma] = \frac{\Lambda^{-1}}{\nu - q -1}$.

I need the marginal distribution of $X$; I think it should be multivariate $t$ but I'm not sure. I know that if I observed $X_1, ..., X_n$ then the posterior of $\mu$ is multivariate $t$ and that the posterior predictive distribution of $X_{n+1}$ is also multivariate $t$. But I'm not so good with the calculations to get the marginal of a single $X$ under the Normal-Inverse-Wishart and I can't find it derived anywhere.

Hopefully someone knows this off the top of their head; I don't need a derivation although a reference would be good.

guy
  • 7,737
  • 1
  • 26
  • 50
  • 1
    You can find this result in the wikipedia article [Conjugate prior](http://en.wikipedia.org/wiki/Conjugate_prior). I hope this helps. –  May 29 '12 at 09:13
  • related: https://stats.stackexchange.com/questions/416164/integrate-out-covariance-matrix-in-normal-wishart-distribution/416197#416197 – Taylor Jul 20 '19 at 03:39

1 Answers1

1

The marginal of a single $X$ is a student t-distribution, referring to this document: https://www.cs.ubc.ca/~murphyk/Papers/bayesGauss.pdf

TomHall
  • 353
  • 2
  • 12