R help - May 2011 - using glmer to fit a mixed-effects model with gamma-distributed response variable

Sub: using glmer to fit a mixed-effects model with gamma-distributed
response variable

Hello,
I'm currently trying to fit a mixed effects model , i.e.:
> burnedmodel1.2<-glmer(gpost.f.crwn.length~lg.shigo.av+dbh+leaf.area+
bark.thick.bh+ht.any+ht.alive+(1|site/transect/plot), family=gaussian,
na.action=na.omit, data=rws30.BL)

If I run this code, I get the error below:

Error: length(f1) == length(f2) is not TRUE
In addition: Warning messages:
1: In plot:(transect:site) :
  numerical expression has 175 elements: only the first used
2: In plot:(transect:site) :
  numerical expression has 175 elements: only the first used

Someone on this forum made the following suggestion,
> rws30.BL$site <- factor(rws30.BL$site)
> rws30.BL$transect <- interaction(rws30.BL$site, rws30.BL$transect, drop
TRUE)
> rws30.BL$plot <- interaction(rws30.BL$site, rws30.BL$transect,
rws30.BL$plot, drop = TRUE)

Which once run allows the above code to run for a gaussian family. However,
if I try to fit a gamma family model (my goal), i.e.
> burnedmodel1.2<-glmer(gpost.f.crwn.length~lg.shigo.av+dbh+leaf.area+
bark.thick.bh+ht.any+ht.alive+(1|site/transect/plot), family=Gamma,
na.action=na.omit, data=rws30.BL)

I get:
Error: no valid set of coefficients has been found: please supply starting
values
In addition: Warning message:
In log(ifelse(y == 0, 1, y/mu)) : NaNs produced

How do I determine what reasonable starting values are? And what's up with
the NaNs?

Thanks for taking a look.


*Ben Caldwell*

PhD Candidate
University of California, Berkeley

	[[alternative HTML version deleted]]

Benjamin Caldwell <btcaldwell <at> berkeley.edu> writes:
> 
> Sub: using glmer to fit a mixed-effects model with gamma-distributed
> response variable
> 
> Hello,
> I'm currently trying to fit a mixed effects model , i.e.:
> 
> >
burnedmodel1.2<-glmer(gpost.f.crwn.length~lg.shigo.av+dbh+leaf.area+
> bark.thick.bh+ht.any+ht.alive+(1|site/transect/plot), family=gaussian,
> na.action=na.omit, data=rws30.BL)
  [stuff about turning site, transect, plot into factors snipped]
> >
burnedmodel1.2<-glmer(gpost.f.crwn.length~lg.shigo.av+dbh+leaf.area+
> bark.thick.bh+ht.any+ht.alive+(1|site/transect/plot), family=Gamma,
> na.action=na.omit, data=rws30.BL)
> 
  Alas, Gamma GLMMs are not yet feasible in glmer -- this is in the
works but I wouldn't hold my breath (try searching the r-sig-mixed-model
archives for this topic). At this point your options are somewhat
limited, to 'build your own model' tools such as WinBUGS or AD Model
Builder (it is conceivable that Gamma GLMMs could be added in the
development version of glmmADMB, but again I wouldn't hold my breath).

  If I may ask, is there a reason you need Gamma GLMMs and not
log-normal GLMMs?  At least qualitatively, the properties of Gamma
and lognormal distributions are reasonably similar (2-parameter families,
domain = non-negative reals, distribution ranges from 'L-shaped' to
approximately normal ...)  You can fit a log-normal GLMM by simply
log-transforming your data (dealing with zeros appropriately) and
fitted a regular, linear mixed model ...

  Further discussions on this topic would probably be better on 
the r-sig-mixed-models list.

  Ben Bolker