I am using a Gaussian model with a conjugate Normal-Inverse-Wishart (NIW) prior, as described here. The advantage of this approach is that the marginal likelihood $p(y)$, which is what I am interested in, is available in closed form.

My problem is that the results seem to be dependent of the NIW hyper-parameters (I have no prior information), with some of the dangers being described here.

As an alternative, I am considering bootstrapping my data in order to obtain $m$ estimate of the mean and covariance. Then I could calculate the marginal likelihood:

$$ p(y) \approx \frac{1}{m} \sum_{i=1}^m p(y|\hat{\mu}_i, \hat{\Sigma}_i). $$

Would this prior be an approximation to an Empirical Bayes prior, something else or just nonsense?

Thank you


Have you considered simply applying a scaling to the covariance matrix, as suggested by Andrew Gelman?

See also this paper on the scaled inverse Wishart.

| cite | improve this answer | |
  • $\begingroup$ That's pretty much what I was doing originally. Then I tried to get a prior that depended only on the data (by bootstrapping), in order to avoid choosing a prior and its hyper-parameters. $\endgroup$ – Matteo Fasiolo Oct 30 '14 at 20:19

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.