[R-sig-ME] R-sig-mixed-models Digest, Vol 95, Issue 6

[Ben Bolker]

Ken Beath <ken.beath at ...> writes:

> 
> nAGQ=0 uses an even more approximate method, so probably isn't advised.
> Looking at your output something has gone seriously wrong. The standard
> errors are all very large and the random effect variance is zero.
> 
> Have you checked whether there is a collinearity problem 
> between your fixed
> effects. Start with a model with all the fixed effects and no random and
> see how that works.

  And/or check for *complete separation*; whenever you have fixed effects
in a GLMM that are large (e.g. |beta|>10) you can suspect that you have
a category that is all-zero or all-one, and bad things will happen as
a result.

Here's your coefficient table:

(Intercept)       -4.185e+01  1.538e+04  -0.003    0.998
Year2007           1.933e+01  1.096e+04   0.002    0.999
Sex               -1.878e+01  1.139e+04  -0.002    0.999
Egg_Volume        -5.620e-03  2.077e-01  -0.027    0.978
Hatch_OrderSecond  1.997e+01  1.079e+04   0.002    0.999
Hatch_OrderThird  -3.482e-01  2.544e+04   0.000    1.000

  So that says that the predicted probability of a 1 in your baseline
condition (first hatch, egg volume zero, sex=0, first year) is
plogis(-41) approx. 1e-18 (!).

  If this does turn out to be the problem, you can solve it
by adding a small prior (e.g. using blmer); see
http://rpubs.com/bbolker/glmmchapter and search for "complete separation".

  I'm just a little surprised that it worked (i.e., ran *and* you
got sensible answers) before ...
 
> On 6 November 2014 13:27, Luciano La Sala <lucianolasala <at> yahoo.com.ar>
> wrote:
> 
> > Dear Ken and Ben,
> >
> > > model.1 <- glmer(Death_2 ~ Year + Sex + Egg_Volume + Hatch_Order +
> > (1|Nest_ID), nAGQ=0, family = binomial, data = surv.2)
> > > summary(model.1)
> >
> > Generalized linear mixed model fit by maximum likelihood (Adaptive
> > Gauss-Hermite Quadrature, nAGQ =  0)
> >  [glmerMod]
> >  Family: binomial  ( logit )
> > Formula: Death_2 ~ Year + Sex + Egg_Volume + Hatch_Order + (1 | Nest_ID)
> >    Data: surv.2
> >
> >      AIC      BIC   logLik deviance df.resid
> >     22.0     44.7     -4.0      8.0      182
> >
> > Scaled residuals:
> >     Min      1Q  Median      3Q     Max
> > -0.2291  0.0000  0.0000  0.0000  4.4713
> >
> > Random effects:
> >  Groups  Name        Variance Std.Dev.
> >  Nest_ID (Intercept) 0        0
> > Number of obs: 189, groups:  Nest_ID, 111
> >
> > Fixed effects:
> >                     Estimate Std. Error z value Pr(>|z|)
> > (Intercept)       -4.185e+01  1.538e+04  -0.003    0.998
> > Year2007           1.933e+01  1.096e+04   0.002    0.999
> > Sex               -1.878e+01  1.139e+04  -0.002    0.999
> > Egg_Volume        -5.620e-03  2.077e-01  -0.027    0.978
> > Hatch_OrderSecond  1.997e+01  1.079e+04   0.002    0.999
> > Hatch_OrderThird  -3.482e-01  2.544e+04   0.000    1.000
> >