In case any other newbie finds this thread in their searching: Someone emailed me off-list and suggested going ahead and progressively ramping up the iterations, and increasing the thin rate and burnin as well. It worked great. All the model diagnostics I can throw at it are looking good. Rafter Sass Ferguson, MS PhD Candidate | Crop Sciences Department University of Illinois in Urbana-Champaign liberationecology.org 518 567 7407 On Thu, Jun 18, 2015 at 4:48 PM, rafter sass ferguson <
liberationecology at gmail.com> wrote:
Hello,
I have a question about a mix of good and bad effective sample sizes. I've
read Jarrod's tutorial, course notes, and paper, and googled around
extensively, and haven't had any luck figuring this out. I'd be grateful
for any guidance.
Working with 721 observations, I fit a 2-level model with three Gaussian
response variables and several fixed effects and interactions. The model
looks good as far as trace plots and gelman diagnostics go.
My concern is with the effective sample size. Eff.samp overall and for
trait:block are quite good (1200, and mostly ~1200 w/ one 920,
respectively). But for trait:ID and residuals (trait:units) eff.samp is
terrible - mostly in the 150-200 range.
I'll post more model details below, but here are my questions:
? Do I need to worry about the poor eff.samp scores if I'm only interested
in the fixed effects?
? If it is a problem, could it be fixed by increasing iterations by a ~10x?
Here is the model I fit -
prior2 <- list(R=list(V=diag(3),nu=3),
G=list(G1=list(V=diag(3), nu=3.02, alpha.mu=rep(0,3),
alpha.V=1000*diag(3)),
G2=list(V=diag(3), nu=3.02, alpha.mu=rep(0,3),
alpha.V=1000*diag(3))
) )
m2b <- MCMCglmm(cbind(professional, relational, practice) ~
-1 + trait +
trait:income + trait:age + trait:ethnicity + trait:gender
+ trait:residence + trait:education + trait:HDI +
trait:Ineq + trait:Enviro + trait:Enviro:gender + trait:Ineq:gender +
trait:Ineq:ethnicity + trait:Ineq:income,
random= ~us(trait):block + us(trait):ID, rcov=
~us(trait):units,
family=rep("gaussian",3),
prior=prior2,
nitt <- 80000, thin <- 25, burnin <- 50000,
data=df_sel,
verbose=TRUE)
Here is the first part of the model summary:
Iterations = 50001:79976
Thinning interval = 25
Sample size = 1200
If a full data set will help, let me know and I'll post one.
Thanks so much for any suggestions!
Warmly,
Rafter
Rafter Sass Ferguson, MS
PhD Candidate | Crop Sciences Department
University of Illinois in Urbana-Champaign
liberationecology.org
518 567 7407