Dear list members,
I fit a random intercept morel with lme4 first, then I tried to fit the same
model with MCMCglmm. I didn't specify the prior and ran 300000 iterations.
The model still didn't converge as the autocorrelation is very high (>0.4).
May I ask (1) do I have to specify my own prior? and (2) what is the
usually/average number of iterations in Bayesian multilevel model? 300000
seems quite large a number. Do I have to rerun the model if the previous one
doesn't converge?
Best with thanks.
>data(Contraception,package="mlmRev")
>library(lme4)
>fm1 <-
glmer(use~urban+age+livch+(1|district),data=Contraception,family=binomial,nAGQ=7)
>MC1 <-
MCMCglmm(use~urban+age+livch,random=~district,data=Contraception,family="categorical",nitt=300000,burnin=240000,thin=100)
>summary(MC1)
Iterations = 240001:299901
Thinning interval = 100
Sample size = 600
DIC: 47.28245
G-structure: ~district
post.mean l-95% CI u-95% CI eff.samp
district 3242 953 6490 132
R-structure: ~units
post.mean l-95% CI u-95% CI eff.samp
units 38025 13640 57500 85.53
Location effects: use ~ urban + age + livch
post.mean l-95% CI u-95% CI eff.samp pMCMC
(Intercept) -197.901 -262.707 -130.966 92.68 <0.002 **
urbanY 86.448 53.374 119.831 131.85 <0.002 **
age -3.139 -5.439 -1.200 166.36 <0.002 **
livch1 128.027 76.737 177.583 126.78 <0.002 **
livch2 159.752 102.289 222.468 109.28 <0.002 **
livch3+ 155.254 92.145 216.916 123.80 <0.002 **
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
>autocorr(MC1$Sol)
--
Wincent Ronggui HUANG
Sociology Department of Fudan University
PhD of City University of Hong Kong
http://asrr.r-forge.r-project.org/rghuang.html
[[alternative HTML version deleted]]