Laplace approximation?

[Aki Vehtari]

http://www.numfocus.org/open-source-projects.html
says about Stan "It features full Bayesian inference with Markov chain Monte Carlo sampling, approximate Bayesian inference with variational inference and the Laplace approximation, and penalized maximum likelihood estimation with optimization. "

I haven't seen the Laplace approximation advertised elsewhere and I think it's not in the manual yet. Can someone tell more?

[Dustin Tran]

This is true, we should be consistent so it’s best to either omit it here or include it everywhere else.

In RStan using optimizing() (and maybe with the optimize command in CmdStan most generally?), you can get draws from a normal approximation; it is centered at the posterior mode and with the inverse Hessian evaluated at the log posterior’s mode as the covariance.

Dustin

> --
> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Bob Carpenter]

I didn't write that blurb!

Laplace approximation around the MLE is in RStan. I
don't think it's made it to CmdStan or PyStan --- we really
want to get Hessians working, but RStan just used
finite diffs on the gradients. That's all there is to
tell, I believe.

- Bob

> On Mar 2, 2016, at 3:36 PM, Aki Vehtari <aki.v...@aalto.fi> wrote:
>