Timeline for Computation of the marginal likelihood from MCMC samples
Current License: CC BY-SA 4.0
21 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| S Dec 23, 2020 at 18:15 | history | suggested | j13r | CC BY-SA 4.0 |
added some more links for those interested in nested sampling
|
| Dec 23, 2020 at 17:35 | review | Suggested edits | |||
| S Dec 23, 2020 at 18:15 | |||||
| May 10, 2018 at 20:03 | history | edited | Xi'an | CC BY-SA 4.0 |
added 54 characters in body
|
| May 10, 2018 at 7:58 | history | edited | Xi'an | CC BY-SA 4.0 |
added 489 characters in body
|
| May 10, 2018 at 7:31 | history | edited | Xi'an | CC BY-SA 4.0 |
added 320 characters in body
|
| Jun 7, 2016 at 11:33 | history | edited | Xi'an | CC BY-SA 3.0 |
added 7 characters in body
|
| Jun 7, 2016 at 8:47 | history | edited | Xi'an | CC BY-SA 3.0 |
deleted 2 characters in body
|
| Jun 7, 2016 at 8:39 | history | edited | Xi'an | CC BY-SA 3.0 |
added 1320 characters in body
|
| May 4, 2016 at 15:03 | comment | added | Florian Hartig | @Xi'an: thanks a lot, I will have a look at both options. | |
| May 4, 2016 at 14:45 | comment | added | Xi'an | @FlorianHartig: Addendum: there is an R package called BayesFactor, developed by Richard Morey, that I have never tried and which foundations I know nothing about... | |
| May 4, 2016 at 14:26 | comment | added | Xi'an | @FlorianHartig: the fact that a generic software like BUGS does not return a generic estimate of $\mathfrak{Z}$ is sort of revealing the extent of the problem. The many solutions that one can find in the specialised literature have not produced a consensus estimate. Hence, my recommendation would be to opt for Geyer's logistic regression solution, which is somewhat insensitive to dimension. | |
| May 4, 2016 at 13:56 | comment | added | Florian Hartig | @Xi'an : very helpful, thanks! Can I ask: of all the mentioned approaches, what would currently be your recommendation if one looks for a general approach that tends to work out of the box (i.e. no tuning / checking required from the user)? I would be especially interested in the case of models with a low (< 50) number of parameters, non-normal posteriors, and strong correlations between parameters. | |
| May 3, 2016 at 9:22 | history | edited | Xi'an | CC BY-SA 3.0 |
added 82 characters in body
|
| May 1, 2016 at 7:40 | comment | added | Xi'an | I added a few more details: the issue in implementing the HPD uniform is to figure out a proper compact approximation for the HPD region. The convex hull of points with high posterior values is (NP?) hard to determine while balls centred at those points may intersect, which creates a secondary normalising constant problem. | |
| May 1, 2016 at 7:38 | history | edited | Xi'an | CC BY-SA 3.0 |
added 607 characters in body
|
| May 1, 2016 at 2:42 | comment | added | lacerbi | Starting from point (1)... I read the relevant articles. The "corrected" harmonic mean estimator seems exactly what I was looking for. It's neat and easy to compute given a MCMC output. So... what's the catch? It doesn't look like the method is being widely used, judging from a quick search on Google Scholar. What are its limitations? (besides the need to identify the HPD regions, which I imagine might become an issue for very complicated posteriors in high dimension). I am definitely going to give it a try -- but I wonder if there is something I need to be wary of. | |
| Apr 30, 2016 at 20:24 | history | edited | Xi'an | CC BY-SA 3.0 |
added 112 characters in body
|
| Apr 30, 2016 at 19:28 | history | edited | Xi'an | CC BY-SA 3.0 |
added 191 characters in body
|
| Apr 30, 2016 at 18:37 | vote | accept | lacerbi | ||
| Apr 30, 2016 at 18:37 | comment | added | lacerbi | (+1) Incredibly rich answer, thank you. This will be useful to me and, I suppose, many other people. It will take me some time to have a look at the various approaches, and then I might come back with specific questions. | |
| Apr 30, 2016 at 18:00 | history | answered | Xi'an | CC BY-SA 3.0 |