1
$\begingroup$

Let $X \sim N(0, \Sigma)$ be a multivariate normal vector, and let our prior for $\Sigma$ be inverse-Wishart: $\Sigma\sim IW(v,V)$. The posterior for $\Sigma$ given $X$ is also inverse-Wishart: $$\Sigma | X = x \sim IW(1 + v, V + x'x).$$

I'm interested in the case in which we observe only some elements of $X$. I.e., we observe $X_c \sim N(0, \Sigma_c)$, where $\Sigma_c$ is a submatrix of $\Sigma$. The posterior density for the full covariance matrix $\Sigma$ given an observation $X_c$ will be $$ p(\Sigma| X_c = x_c) \propto |\Sigma_c|^{-\frac{1}{2}} \exp\left(-\frac{1}{2} x_c' \Sigma_c^{-1} x_c\right)f(\Sigma), $$ where $f(\Sigma)$ is the prior density of $\Sigma$.

Is there a prior on $Σ$ such that this posterior is conjugate for any set of observed elements of $X$?

$\endgroup$

1 Answer 1

1
$\begingroup$

While I haven't been able to find a conjugate prior, I believe the following MCMC algorithm will suffice:

In each iteration $r$:
($r.1$) draw the missing elements ${X_{-c}}^r$ given $X_c, Σ^{r-1}$,
($r.2$) draw $Σ^r$ given $X_c, {X_{-c}}^r$.

Step ($r.1$) just requires drawing from a Normal distribution while step ($r.2$) can draw from an Inverse-Wishart as described in the original question.

I believe that $Σ^r$ should converge in distribution to $Σ | X_c$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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