[ADMB Users] Transformation of mcmc values

Ben Bolker bbolker at gmail.com
Tue Mar 20 09:45:13 PDT 2012

On 12-03-20 12:02 PM, Håkon Holand wrote:

> I was wondering if anyone knows a way of transforming values obtained from a HPD of mcmc runs of a model?
> I am running models with a binary response variable (0/1) (family=binomial) and using individual identity as a random variable. I am currently using R 2.14.1 (64-bit) and glmmadmb v.
> The intervals looks like this:

>> head(HPDinterval(m1))
>               lower      upper
> beta.1 -487.3376278 -389.89072
> beta.2   88.3658266  167.10142
> beta.3 -106.7078281  -46.53788
> beta.4    1.2919346   34.92751
> beta.5   -0.8259944   50.58856
> beta.6  -49.8802241  -12.91753

> First of all: I know that these values are from the parameters fitted internally, using an orthogonalized version of the original design matrix, not the original coefficients.

> The question is:  (if possible) How can I transform these values to, for example, Logit? I would like to have my estimated value and its limits on the same scale at least (and hopefully a "reader friendly" scale).

http://markmail.org/message/ga5lh6hbyhh2iqht should help, plus info in
the latest (upcoming?) version of the package vignette.

  You do need to watch out for back-transforming confidence intervals
on the logit scale, though.  Back transforming the *parameters* (i.e.
the differences in the logit per unit of each predictor) to the
probability scale often doesn't make sense. Usually, only
back-transforming *predictions* to the probability scale makes sense.
This is a fundamental conceptual problem that applies to GLMs on the
logit scale, not just to GLMMs ...


More information about the Users mailing list