[R] glmer vs glmmadmb

[Ben Bolker]

kenny xu <kennyxu1983 <at> hotmail.com> writes:

> 
> Dear All
>
> I have fitted the following glmm:
> 
> cmai ~ time.f * intrv.f + (1 | nhome.f/Res_Code.f)
> 
> with poisson distribution, using both glmer and glmmadmb.
> 
> But the estimation for the fixed and random effects were different, i.e.

  This is a surprising set of differences.  I'm going to suggest
you send follow-ups to r-sig-mixed-models at r-project.org, which is
specialized for mixed models

> > summary(lmer.AGGREG.cmai.out3)
> 
> Call:
> glmmadmb(formula = cmai ~ time.f * intrv.f + (1 | nhome.f/Res_Code.f), 
>     data = beam.AGGREG.cmai.long, family = "poisson", link = "log", 
>     zeroInflation = F, admb.opts = admbControl(impSamp = 0, run = F), 
>     save.dir = "tmp")

   Is there a particular reason you're using 'run=FALSE'?  This specification
will tell glmmADMB not to run the model, but to collect the results of
a previous run from the working directory -- not necessarily wrong,
but very easy to make a mistake this way and pick up the results
from a model run with a *different* specification (which might???
be what happened here)  (Also, just as a matter of practice, it's
strongly advised to use FALSE instead of F, just in case someone
decided to assign a value to 'F' ...)

> AIC: 1032.2 
> 
> Coefficients:
>                  Estimate Std. Error z value Pr(>|z|)
> (Intercept)         0.542      3.105    0.17     0.86
> time.f2             0.104      5.177    0.02     0.98
> time.f3            -0.526      3.230   -0.16     0.87
> intrv.f1            0.929      2.712    0.34     0.73
> time.f2:intrv.f1   -0.416      5.302   -0.08     0.94
> time.f3:intrv.f1    0.177      3.261    0.05     0.96
> 
> Number of observations: total=1032, nhome.f=35, nhome.f:Res_Code.f=344 
> Random effect variance(s):
> Group=nhome.f
>             Variance StdDev
> (Intercept)   0.7118 0.8437
> Group=nhome.f:Res_Code.f
>             Variance StdDev
> (Intercept)    1.454  1.206
> Log-likelihood: -508.108 
> 
> > summary(lmer.AGGREG.cmai.out2)
> Generalized linear mixed model fit by the Laplace approximation 
> Formula: cmai ~ time.f * intrv.f + (1 | nhome.f/Res_Code.f) 
>    Data: beam.AGGREG.cmai.long 
>   AIC  BIC logLik deviance
>  1835 1874 -909.5     1819
> Random effects:
>  Groups             Name        Variance Std.Dev.
>  Res_Code.f:nhome.f (Intercept) 0.040125 0.20031 
>  nhome.f            (Intercept) 0.033702 0.18358 
> Number of obs: 1032, groups: Res_Code.f:nhome.f, 344; nhome.f, 35
> 
> Fixed effects:
>                  Estimate Std. Error z value Pr(>|z|)    
> (Intercept)       3.62040    0.04749   76.23   <2e-16 ***
> time.f2          -0.01964    0.01706   -1.15   0.2496    
> time.f3           0.01643    0.01691    0.97   0.3310    
> intrv.f1          0.07540    0.06819    1.11   0.2689    
> time.f2:intrv.f1  0.02148    0.02395    0.90   0.3698    
> time.f3:intrv.f1 -0.04835    0.02394   -2.02   0.0435 *  

  Otherwise I'm stumped.  The numbers of observations etc. etc.
seem consistent. It's hard to compare AIC/log-likelihood between
glmmADMB and glmer because (at present) they use different
additive offsets ...

   You could send me the data if it's not too sensitive.

  Ben Bolker