Skip to content
Prev 15410 / 20628 Next

persistant autocorrelation in binomial MCMCglmm

Hi Jarrod,

I have an update on the model performance with a probit link - the
autocorrelation is much better behaved with the "threshold" family. All ACF
values are <0.1 for iteration and thinning #s that were resulting in
autocorrelation using the logit link.

Looking at the latent variables, though, a lot of the distributions
included values below -7 (lowest was -10). All of the means where within
the -7 to 7 range though, because only a few estimates per observation
tended to reach very low negative values. Are *any* estimates outside the
range considered problematic?

I noticed on the forum a post where someone else had this issue (
https://stat.ethz.ch/pipermail/r-sig-mixed-models/2012q3/019067.html) and I
also tried the chi square prior you suggested for that problem (V=1,
nu=1000, alpha.mu=0, alpha.V=1) but the result was the same.

In terms of the data and system, I would expect an extremely low, near zero
probability of interaction for some of these dyads because they are not
using similar areas and so are not physically able to interact. Is this
signal perhaps too strong? If my goal is to weed out these improbable
interactions, though, will the model not serve this purpose?

Many thanks,

Christina

Christina M. Aiello
Biologist- U.S. Geological Survey
Las Vegas Field Station
160 N. Stephanie St.
Henderson, NV 89074
(702) 481-3957
caiello at usgs.gov
On Thu, May 4, 2017 at 10:17 AM, Aiello, Christina <caiello at usgs.gov> wrote: