Skip to content

glmm question

4 messages · Ken Beath, Joaquín Aldabe

Original
Joaquín Aldabe
# 
Hello, this is Joaqu?n Aldabe from Uruguay. I?m trying to model shorebird
counts (Buff breasted Sandpiper, BBSA) with glmm (using lme4 package),
using continuous variables (grass height, field area, forest cover) and one
factor variable (presence/absence of other shorebird species: American
Golden Plover, AMGP). I sampled 19 fields during three years in December.
I?m interested in identifying predictors correlated with BBSA counts. I
used Year as a random effect as I?m not interested in Year as a fix effect
and because fields were counted three times (pseudoreplication).

The model doesn?t converge, and the output showed that the factorial
variable has not a significant effect. This is weird as in every field I
observed the Buff breasted Sandpiper I also observed the other
species. When I take AMGP out, the model runs ok.

This is the model I?m trying to run:

mysub3.3<-glmer(BBSA~Grass_height+Field_area+Field_enclosure_700m+Grass_height*Field_enclosure_700m+fAMGP+(1|fYear),family="poisson",
data=mysub3.2)

continuous variables were scaled.

I can send de data frame if somebody is interested.

Thanks in advanced for helping me on my master thesis.

Cheers,

Joaqu?n.
Ken Beath
# 
Having a random effect with only 3 levels is not recommended, it usually
gives problems fitting. There are also some philosophical questions about
its use as a random effect.

A random effect for field is reasonable, but you may be fitting too many
parameters. With only 57 observations it is easy to overfit the models, and
a standard linear model may be all that is necessary.
On 14 May 2015 at 07:15, Joaqu?n Aldabe <joaquin.aldabe at gmail.com> wrote:

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Joaquín Aldabe
    

    

      
      #
    

  


  

      
Thanks a lot Ken for your response. I had decided to use mix model with
random effects because I have repeated measures in each field (one per
year). If I perform a glm, how can I manage the pseudoreplication?
(repeated counts on the same field)

I have one data per field per year. So, no hierarchical or nested structure
of the data. Can I use field as a random effect anyway?

Thanks a lot for your help.

Cheers,
Joaqu?n.

2015-05-13 19:17 GMT-03:00 Ken Beath <ken.beath at mq.edu.au>:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Ken Beath
    

    

      
      #
    

  


  

      
I think you would be better off to start with a mixed effects model with a
random effect for site and a fixed effect for year.

All a random effect is doing is to model the correlation between responses
allowing for other variables, in other words the residuals, by conditioning
on a variable that is unobserved. For the years there are only 3 so it is
just as easy or easier to model them using a known variable, the actual
year. It is also difficult to think of them as a random sample of years.
Now for the sites you would expect the same, that the 3 measurements within
a site, representing the 3 years would be correlated. Now it is reasonable
to model them using a random effect, as otherwise there would need to be a
fixed effect for each site, a large number of parameters. It is possible
that this random effect has variance zero then the model reverts to a
standard glm.

        
On 14 May 2015 at 22:55, Joaqu?n Aldabe <joaquin.aldabe at gmail.com> wrote: