[R-sig-ME] Different results and warning messages from an identical glmer model at different times.

[Hsu, B.]

Hello,

About October and November last year, I used the glmer function in lme4 to
analyze some data of my experiment. I made two models to analyze the
probably of embryo development and hatching success of a species of birds,
as below:

glmer(developed ~ Treatment + Layingorder + Eggmass + Layingdate + (1|Nest
of egg laying) + (1|Nest of incubation), data, family="binomial")
glmer(hatching ~ Treatment + Layingorder + Eggmass + Layingdate + (1|Nest
of egg laying) + (1|Nest of incubation), data, family="binomial")

Both models worked fine at that moment.

However, now I am re-running these models but the results become different,
and the models give some warning messages which did not come out before. I
am sure that the data are not modified since last time and the code is
identical. I also checked the version of lme4, which is 1.1-7, and this
version seems to be published on 2014-7-19. Therefore, the version I am
using now is the same as last time. This makes me so puzzled why I got
different results and why I got warning messages that I did not have at
last time.

After some work of google and reading, I still could not find any clue to
solve this problem. Given the fact that I ran the same model twice with the
same version of lme4 but at different times, would it be due to any update
in lme4?

The warning messages I got are copied below:

for the model of embryo development:

Warning messages:
1: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model failed to converge with max|grad| = 4.04423 (tol = 0.001, component
3)
2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model failed to converge: degenerate  Hessian with 1 negative eigenvalues

for the model of hatching success:

Warning messages:
1: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model failed to converge with max|grad| = 0.970515 (tol = 0.001,
component 3)
2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model is nearly unidentifiable: very large eigenvalue
 - Rescale variables?;Model is nearly unidentifiable: large eigenvalue ratio
 - Rescale variables?

Thank you very much in advance,
Bin-Yan Hsu

-- 
Bin-Yan Hsu*, *PhD candidate,
Behavioural Biology,
Groningen Institute for Evolutionary Life Sciences,
Faculty of Mathematics and Natural Sciences,
University of Groningen

	[[alternative HTML version deleted]]