On 3/27/2009 7:49 AM, imicola wrote:
Hi,
Im carrying out some Bayesian analysis using a binomial response variable
(proportion: 0 to 1), but most of my observations have a value of 0 and many
have very small values (i.e. 0.001). I'm having troubles getting my MCMC
algorithm to converge, so I have decided to try normalising my response
variable to see if this helps.
It seems to me that the problem in a situation like this is with the
algorithm, not with the data. Can't you modify it to get better
convergence? For example, set your target to be the square root of your
posterior (or some other power between 0 and 1); this is more diffuse,
so it's easier to sample from. Then use importance sampling to reweight
the sample.
Duncan Murdoch
I want it to stay between 0 and 1 but to have a larger range of values, or
just for them all to be slightly higher.
Does anyone know the best way to acheive this? I could just add a value to
each observation (say 10 to increase the proportion a bit, but ensuring it
would still be between 0 and 1) - would that be ok? Or is there a better
way to stretch the values up?
Sorry - i know its not really an R specific question, but I have never found
a forum with as many stats litterate people as this one :-)
Cheers - any advice much appreciated!
nicola
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
