search

Conte-Ecology / conteStreamTemperature

Package for cleaning and analyzing stream daily stream temperature
MIT License
1 stars 1 forks source link

scale wishart prior #33

Closed djhocking closed 9 years ago

djhocking commented 9 years ago

The inverse wishart as a prior on a multivariate normal can have problem of constraining the combination of covariances resulting in accidentally informative priors on the standard deviations of the random effects. Gelman and Hill recommend the scaled inverse wishart and give an example of implementation in BUGS. I will try that.

djhocking commented 9 years ago

I added the scaled inverse wishart but it slowed convergence from 100s of iterations to 10,000s of iterations. The model took 2 weeks to run but did converge and looks well mixed. It was surprising that it slowed things down that much, especially since Gelman suggested it should speed up convergence. I did save the raw tau and xi values so I could use more sensible starting values next time.

djhocking commented 9 years ago

More info on scaled inverse wishart and other options for priors on multivariate normals

http://www.themattsimpson.com/2012/08/20/prior-distributions-for-covariance-matrices-the-scaled-inverse-wishart-prior/

https://dahtah.wordpress.com/2012/03/07/why-an-inverse-wishart-prior-may-not-be-such-a-good-idea/

I contacted the authors of one of the papers referenced and they were unable to get the separation method to work in JAGS.