[R-sig-ME] Considerable discrepancies between fixed and random effect estimates of lme4 (glmer) and glmmADMB

[Adam Smith]

> > I hope I'm not overlooking something elementary here, but estimated
> > fixed and random effects are considerably different from the
> > following Poisson model in lme4 and glmmADMB.  The fixed effects
> > seem to differ most considerably...  Thanks for any thoughts...
> > Adam Smith
>
> > > sessionInfo()
> > R version 2.14.1 (2011-12-22)
> > Platform: x86_64-pc-mingw32/x64 (64-bit)
> >
> > (SNIP)
> >
> > other attached packages:
> > [1] glmmADMB_0.7.2.5   lme4_0.999375-42   Matrix_1.0-3
> bbmle_1.0.4.1      numDeriv_2010.11-1
> > [6] lattice_0.20-0     R2admb_0.7.5       MASS_7.3-16
>
>
> > > str(cons09) # The dataset
> > 'data.frame':   394 obs. of  15 variables:
> >  $ plot       : Factor w/ 16 levels "n_10","n_2","n_3",..:
> > 2 2 2 2 2 2 2 2 2 2 ...
> >  $ plot_trt   : Factor w/ 32 levels
> > "cont_n_10","cont_n_2",..: 2 2 2 2 2 2 2 2 2 2 ...
> >  $ geog       : Factor w/ 2 levels "n","s": 1 1 1 1 1 1 1 1 1 1 ...
> >  $ trt        : Factor w/ 2 levels "cont","trt": 1 1 1 1 1 1 1 1 1 1 ...
> >  $ count      : Factor w/ 14 levels "1","2","3","4",..:
> > 1 2 3 4 5 6 7 8 9 10 ...
> >  $ total      : int  341 326 257 244 185 141 128 121 115 84 ...
> >  $ cons       : int  12 52 8 57 36 8 0 1 20 27 ...
> >  $ dt         : int  4 3 3 3 3 3 3 3 3 3 ...
> >  $ obs        : Factor w/ 394 levels "1","2","3","4",..:
> > 1 2 3 4 5 6 7 8 9 10 ...
> >  $ logtotal   : num  5.83 5.79 5.55 5.5 5.22 ...
> >  $ logdt      : num  1.39 1.1 1.1 1.1 1.1 ...
>
>
> > # The models
> > > Poiss <- glmmadmb(cons ~ count + geog + trt + count:trt +
> > count:geog + trt:geog + offset(logtotal) +
> >         offset(logdt) + (1|plot) + (1|plot_trt),
> > zeroInflation=FALSE, family="poisson", data=cons09)
> >
> > > P_glmer <- glmer(cons ~ count + geog + trt +
> > count:trt + count:geog + trt:geog + offset(log(total)) +
> >       offset(log(dt)) + (1|plot) +
> > (1|plot_trt), family="poisson", data=cons09)
>
> By the way, you can specify these fixed effects
> more compactly as (count+geog+trt)^2 ...

Indeed, I was being explicit for explicitness' sake...

>
> > > fixef(Poiss)
>
> > 0.77333 0.58885 -0.21547 0.79101 0.24924 -0.02565 0.36004
> > -0.94852 0.39442 0.29032 0.13658 -0.39564
> > -1.37470 -0.88496 -0.53259 -0.25434
> > -0.41033 1.18840 0.25678 0.69734
> > 0.59823 1.11430 -0.17330 -0.15364
> > 0.98442 0.51533 -0.99431 -1.63920 -0.40455 -1.55670 0.15891
> > 0.35378 0.45748 0.92293 0.78373 1.05440
> > 0.39598 0.32019 1.41600
> > 0.86656 2.45790 1.07350 -0.31149
>
> > # After detaching glmmADMB and lme4, then
> > re-requiring lme4 to avoid masking of lme4's fixef function
> > > fixef(P_glmer)
> > -5.00326871 0.69554781 0.17105622 1.29467717
> > 0.97268949 0.87707266 1.39399967 0.22923370
> > 1.61829807 1.84546906 1.99506279 1.37859852
> > 0.53987085 1.13282524 -0.12901214 0.07021103
> > -0.52669841 0.87378323 0.05397134 0.46901053
> > 0.39863837 0.94486828 -0.18146287 -0.20340655
> > 0.77866847 0.49935778 -0.65101999 -1.05381279
> > -0.05290984 -1.54103716 -0.03887633 0.04570463 0.02076874
> > 0.47883964 0.23057821 0.49408747
> > -0.15116087 -0.32683907 0.50098187 0.14670125
> > 1.63961096 0.40944384 -0.22555478
> >
> > I juxtapose ranef() estimates here for comparison's sake...
> >
> [snip]
>
> There's nothing obviously wrong here. It's not a full solution,
> but I wonder how wide the confidence intervals are ... if they
> are very wide, then the practical answer is that these are poorly
> determined estimates. You're probably overfitting the model --
> 2 random effects plus 43 fixed-effect coefficients is
> quite a lot for 394 observations (the general rule of thumb
> is N/(# params)>10), especially if the Poisson data are sparse
> (although they don't look that way from the first few
> values listed in str() -- is there anyway you can allow
> 'count' to be continuous, or ordinal, rather than insisting
> on it being categorical?) But it would be best to try to answer
> more definitively ... can you send data?

I started with a general specification of count expecting to model 
it with an additive term when I start comparing fixed effects.  I 
suppose I could model it as a lower order polynomial initially... 
Doing so should drastically reduce the number of fixed effects in 
the model.  I suppose I should center it before creating the
polynomial terms...

However, I'll still send the data off-list.  

Thanks for looking this over.

>
> Ben Bolker
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models