Skip to content

problems with 'false convergence' in lmer

4 messages · Andrew J Tyre, Toni Hernandez-Matias

Original
Andrew J Tyre
# 
Toni,

I assume that your predictor variables vary by territory, and that your 
random effect is applied to the intercept, like this

occupancy_ij ~ predictor1_i  + predictor2_i + (1|territoryID_i) + 
(1|year_j)

in that case, the territoryID random effect is "competing" to explain 
variation in occupancy between territories with predictors1 and 2. Hence 
the failure to converge. 

A better question to ask yourself is this: is the probability that a 
territory is occupied this year different depending on whether the 
territory was occupied last year? If yes, then you have AUTOCORRELATION, 
and that has to be handled differently - a random effect on territoryID 
influences the probability of the territory being occupied, as do the 
predictor variables, but that model still assumes that occupancy is 
independent from year to year. I do not know how to accomadate 
autocorrelation in a binomial model in lmer; in general, it isn't 
something that is easy to do. 

1) You cannot use the estimates from lmer if it hasn't converged, in my 
opinion. 

2) not sure ... looking forward to an answer from greater gurus. 

hth,

Drew Tyre

School of Natural Resources
University of Nebraska-Lincoln
416 Hardin Hall, East Campus
3310 Holdrege Street
Lincoln, NE 68583-0974

phone: +1 402 472 4054 
fax: +1 402 472 2946
email: atyre2 at unl.edu
http://snr.unl.edu/tyre
http://aminpractice.blogspot.com



Toni Hernandez-Matias <ahmatias at gmail.com> 
Sent by: r-sig-mixed-models-bounces at r-project.org
06/22/2009 09:54 AM

To
R-SIG-Mixed-Models at r-project.org
cc

Subject
[R-sig-ME] problems with 'false convergence' in lmer






Dear all,

I am analyzing a data set with the 'lmer' function (lme4_0.999375-28).
The dependent variable is binary: occupation status of a breeding site by 
a
bird species (territory ocupied vs. non-ocupied).
There are clustered observations: observations of the same territory for
several years and observations of different territories in the same year. 
So
I considered two random factors: territory identity and year.
I considered several predictors which are standardized.
I am using aic values to obtain the final model.

The problem is that the program generates the warning message in most of 
the
fitted models:
In mer_finalize(ans) : false convergence (8)

This probably happens because many territories are either occupied or
non-occupied for many or all years in the study. [there are no problems 
when
I include the random factor year in the formula, and I do not include the
random factor territory identity]

I added the  option 'verbose=TRUE', and  I found that the printed value of
deviance does not match with that saved by the model object (I extract it
using 'attr(summary(nameofthemodel),"AIC")$AIC').

My questions are:

(1) are the results (estimated parameters and se) of the fitted model
correct? Can I use them?

(2) how can I extract the deviance value of the printed table given by the
'verbose' option? I mean extract, as a R object. I need this because I am
performing a large set of models and I consider those with better aic.

Thank you very much in advance,


Toni Hernandez