Glmer2stan

Glmer2stan is an interesting software that translates glmer (lme4) syntax into Stan. The software is still at eary stage but looks very promising.

A question

I posted a question to the R-Mixed-Model mailing list but have not had any responses yet:
---------------------------------------------------------------------------------------

Dear All,

I wonder if anybody has tried to make glmmBUGS work with JAGS. My
attempt was not successful.

Here is the simple example I copied from the glmmBUGS Vignettes:

------------------------
library(MASS)
data(bacteria)

bacterianew <- bacteria
bacterianew$yInt = as.integer(bacterianew$y == "y")

levels(bacterianew$trt) <- c("placebo", "drug", "drugplus")

library(glmmBUGS)

bacrag <- glmmBUGS(formula = yInt ~ trt + week, data=bacterianew,
effects = "ID",
                   modelFile="model.bug", family="bernoulli")

names(bacrag$ragged)

source("getInits.R")

startingValues = bacrag$startingValues
------------------------

With WinBUGS, it runs well:

------------------------
library(R2WinBUGS)

bacResult1 = bugs(bacrag$ragged, getInits, model.file="model.bug", n.chain=3,
  n.iter=2000, n.burnin=100, parameters=names(getInits()), n.thin=10)
------------------------

But with JAGS, I got error message "Error in FUN(50L[[1L]], ...) :
invalid first argument"

------------------------
library(R2jags)

jags.parms=names(getInits())

bacResult = jags(data=bacrag$ragged, n.chain=3, n.iter=2000,
model.file=model.bug)
------------------------

I hope somebody can help me figuring out how to make this work. Many thanks.

Best,
Shige

---------------------------------------------------------------------------------------

Here is the answer from Jens Åströ
---------------------------------------------------------------------------------------
Hi!

Not an expert by any means, but I got it to run by doing this:

1) change inprod2 to inprod in the model file (JAGS does not have
inprod2 function)

2) change ~dflat() to other uninformative prior, e.g.
~dnorm(0.0,1.0E-6) (Jags does not have dflat distribution)

3) specify/compile the model with e.g.
bac.jags<-jags.model("model.bug",data=bacrag$ragged,n.chains=4)

4) update it, update(bac.jags,1000)

5) Collect coda samples,
bacResult<-coda.samples(temp,names(getInits()),n.iter=10000,thin=10)

This seemed to converge well, (gelman.diag(bacResult))

Good luck
----------------------------------------------------------------------------------------

It works.

GLMM revisted

A short while ago, I reported some discrepancies between the results produced by "lme4" and other R packages as well as Stata. Today I upgraded to the most recent version of "lme4a" and re-ran my model. The error of false convergence disappeared and the new results are very much in line with other packages.

I am glad that the problem eventually got sorted out.

For this version, the adaptive GH quadrature has not been implemented yet; so the only option to estimate a GLMM is laplace transformation.

An interesting paper

Ben Bolker has an interesting paper (outline of a paper) comparing different approaches to estimate GLMM in R environment, which is very helpful to what I am doing right now.

The paper pointed out the following options to fit GLMM using R:

1. glmer
2. glmmML
3. glmm (from the repeated package)
4. glmmADMB
5. MCMCglmm
6. glmmBUGS
7. glmmPQL
8. BUGS (through R2WinBUGS)
9. glmmAK

And I would like to add one more, npmlreg.

I am not aware of the glmmAK package before. From the first glance, it seems to be very promising in the sense that it seems to allow non-Gaussian random effect in a Bayesian framework, something similar to what npmlreg does with ML method.

============== edited on March 3 ====================

DPpackage is another package that can estimate GLMM in a Bayesian framework.

============== edited on March 5 ====================

ASReml/ASReml-R is another choice. It is not free software, but it does seem to have some unique strengths. Maybe I should download a demo copy and try it myself.

============== edited on March 29 ===================

hglm is another possibility. 

Generalized linear mixed effect model problem

I am trying to compare cohort difference in infant mortality using generalized linear mixed model. I first estimated the model in Stata:

xi:xtlogit inftmort i.cohort, i(code)

which converged nicely:


Fitting comparison model:


Iteration 0:   log likelihood = -1754.4476
Iteration 1:   log likelihood = -1749.3366
Iteration 2:   log likelihood = -1749.2491
Iteration 3:   log likelihood = -1749.2491


Fitting full model:


tau =  0.0     log likelihood = -1749.2491
tau =  0.1     log likelihood = -1743.8418
tau =  0.2     log likelihood = -1739.0769
tau =  0.3     log likelihood = -1736.4914
tau =  0.4     log likelihood = -1739.5415


Iteration 0:   log likelihood = -1736.4914  
Iteration 1:   log likelihood = -1722.6629  
Iteration 2:   log likelihood = -1694.9114  
Iteration 3:   log likelihood = -1694.6509  
Iteration 4:   log likelihood =  -1694.649  
Iteration 5:   log likelihood =  -1694.649  


Random-effects logistic regression              Number of obs      =     21694
Group variable: code                            Number of groups   =     10789


Random effects u_i ~ Gaussian                   Obs per group: min =         1
                                                               avg =       2.0
                                                               max =         9


                                                Wald chi2(2)       =      8.05
Log likelihood  =  -1694.649                    Prob > chi2        =    0.0178


------------------------------------------------------------------------------
    inftmort |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
  _Icohort_2 |  -.5246846   .1850328    -2.84   0.005    -.8873422   -.1620269
  _Icohort_3 |  -.1424331    .140369    -1.01   0.310    -.4175513     .132685
       _cons |  -5.214642   .1839703   -28.35   0.000    -5.575217   -4.854067
-------------+----------------------------------------------------------------
    /lnsig2u |   .9232684   .1416214                      .6456956    1.200841
-------------+----------------------------------------------------------------
     sigma_u |   1.586665   .1123528                      1.381055    1.822885
         rho |   .4335015   .0347791                      .3669899    .5024984
------------------------------------------------------------------------------
Likelihood-ratio test of rho=0: chibar2(01) =   109.20 Prob >= chibar2 = 0.000

Then I tried the same model using lme4:

m2 <- glmer(inftmort ~ (1|code) + as.factor(cohort), family=binomial, data=d)

I got:


Warning message:
In mer_finalize(ans) : false convergence (8)

And the results are quite different from what I got from Stata. I tried to estimate the model using "glmmML" and also got into trouble. This time the error message is:


Warning message:
In glmmML.fit(X, Y, weights, cluster.weights, start.coef, start.sigma,  :
  Hessian non-positive definite. No variance!

I then tried MCMCglmm:


prior=list(R=list(V=1, nu=0, fix=1), G=list(G1=list(V=1, nu=0)))
m2 <- MCMCglmm(inftmort ~ 1 + as.factor(cohort), random=~code, family = "categorical", data=d, prior=prior)

It seems to be working and produced estimates that are comparable to what Stata produced (not identical, of course):


Iterations = 3001:12991
Thinning interval = 10 
Number of chains = 1 
Sample size per chain = 1000 


1. Empirical mean and standard deviation for each variable,
   plus standard error of the mean:


                      Mean     SD Naive SE Time-series SE
(Intercept)        -5.7145 0.1805 0.005708        0.02295
as.factor(cohort)2 -0.5633 0.1788 0.005653        0.02194
as.factor(cohort)3 -0.1888 0.1471 0.004653        0.01912


2. Quantiles for each variable:


                      2.5%     25%     50%      75%   97.5%
(Intercept)        -6.0464 -5.8324 -5.7251 -5.61120 -5.2817
as.factor(cohort)2 -0.8977 -0.6985 -0.5551 -0.43084 -0.2371
as.factor(cohort)3 -0.4614 -0.2947 -0.1867 -0.09644  0.1232 

This is puzzling.