default optimizer in lmer 1.1-7?

Thu, Nov 6, 2014 8:11 AM

[cc'ing to r-sig-mixed-models]

You can see the default optimizer for GLMMs by running

 > glmerControl()$optimizer
[1] "bobyqa"      "Nelder_Mead"

This means that bobyqa is used for the first (nAGQ=0) stage of
optimization, whereas Nelder-Mead is used for the second (nAGQ>0).  If
optimizer is a length-1 vector, the same optimizer is used for both stages
of optimization (I *think* this is explained in the documentation, but it
might not be ...)

If you read the NEWS file <
http://cran.r-project.org/web/packages/lme4/news.html> slightly more
carefully, you'll see that the default was only changed for `lmer` fits,
not `glmer` fits -- from 1.1-4,

* The default optimizer for lmer fits has been switched from "Nelder_Mead"
to "bobyqa" because we have generally found the latter to be more reliable.
To switch back to the old behaviour, use
control=lmerControl(optimizer="Nelder_Mead").

  We are very seriously considering making bobyqa, or nloptwrap, the
default for GLMMs as well (for both stages of optimization) in the next
release.

 I haven't looked at your results in too much detail otherwise -- it's a
little bit puzzling that all the AICs seem to be about the same, but I was
in a hurry (and I want to send this off before the problem gets buried in
my inbox).


On Thu, Nov 6, 2014 at 10:16 AM, Jordi Moya Lara?o <jordi at eeza.csic.es>
wrote:

 Dear Ben,
We have been going crazy about replicating previous analyses with glmer
and I think that what may be happening is that when you copiled the last
version you left in unadvertedly "Nelder_Mead" by default instead of
"bobyqa" as it is claimed to be done in the last version of the manual. Is
that a possibility? My analyses give the results by default as those
obtained if you were using "Nelder_Mead". I am also wondering whether
Nelder_Mead results should be trusted whatsoever.

Here are my codes and results (the results using "bobyqa" are the once
qualitatively very similar to those using previous versions of lme4):

TELA_MANIP1<-glmer(ABUNDANCE~RAIN*TRAT+(1|SITE)+(1|caso),data=tela2013,poisson)

summary(TELA_MANIP1)

Generalized linear mixed model fit by maximum likelihood (Laplace
Approximation) ['glmerMod']
 Family: poisson  ( log )
Formula: ABUNDANCE ~ RAIN * TRAT + (1 | SITE) + (1 | caso)
   Data: tela2013

     AIC      BIC   logLik deviance df.resid
   265.8    282.2   -124.9    249.8       49

Scaled residuals:
     Min       1Q   Median       3Q      Max
-1.67352 -0.56382  0.04347  0.63915  1.61069

Random effects:
 Groups Name        Variance  Std.Dev.
 caso   (Intercept) 0.0815360 0.28555
 SITE   (Intercept) 0.0001355 0.01164
Number of obs: 57, groups:  caso, 57; SITE, 7

Fixed effects:
                          Estimate Std. Error z value Pr(>|z|)
(Intercept)               1.515432   0.004537   334.0  < 2e-16 ***
RAINLOW                   0.013886   0.004533     3.1  0.00219 **
TRATNO_PREDATORS         -0.621030   0.004532  -137.0  < 2e-16 ***
TRATPREDATORS             0.238772   0.004534    52.7  < 2e-16 ***
RAINLOW:TRATNO_PREDATORS -0.251747   0.004531   -55.6  < 2e-16 ***
RAINLOW:TRATPREDATORS     0.046884   0.004532    10.3  < 2e-16 ***
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

Correlation of Fixed Effects:
              (Intr) RAINLOW TRATNO TRATPR RAINLOW:TRATN
RAINLOW       0.000
TRATNO_PRED   0.000  0.000
TRATPREDATO   0.000  0.000   0.001
RAINLOW:TRATN 0.000  0.000   0.000  0.000
RAINLOW:TRATP 0.000  0.000   0.001  0.000  0.000

library(car)
Anova(TELA_MANIP1,type="III",test="Chisq")

Analysis of Deviance Table (Type III Wald chisquare tests)

Response: ABUNDANCE
                 Chisq Df Pr(>Chisq)
(Intercept) 1.1158e+05  1    < 2e-16 ***
RAIN        9.3829e+00  1    0.00219 **
TRAT        2.1559e+04  2    < 2e-16 ***
RAIN:TRAT   3.1937e+03  2    < 2e-16 ***
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

TELA_MANIP1<-glmer(ABUNDANCE~RAIN*TRAT+(1|SITE)+(1|caso),data=tela2013,poisson,glmerControl(optimizer
= c("bobyqa")))

summary(TELA_MANIP1)

Generalized linear mixed model fit by maximum likelihood (Laplace
Approximation) ['glmerMod']
 Family: poisson  ( log )
Formula: ABUNDANCE ~ RAIN * TRAT + (1 | SITE) + (1 | caso)
   Data: tela2013
Control: glmerControl(optimizer = c("bobyqa"))

     AIC      BIC   logLik deviance df.resid
   265.8    282.2   -124.9    249.8       49

Scaled residuals:
     Min       1Q   Median       3Q      Max
-1.67891 -0.56610  0.04277  0.63769  1.61159

Random effects:
 Groups Name        Variance Std.Dev.
 caso   (Intercept) 0.08089  0.2844
 SITE   (Intercept) 0.00000  0.0000
Number of obs: 57, groups:  caso, 57; SITE, 7

Fixed effects:
                         Estimate Std. Error z value Pr(>|z|)
(Intercept)               1.51734    0.19288   7.867 3.64e-15 ***
RAINLOW                   0.01200    0.30771   0.039   0.9689
TRATNO_PREDATORS         -0.62228    0.27051  -2.300   0.0214 *
TRATPREDATORS             0.23893    0.23320   1.025   0.3056
RAINLOW:TRATNO_PREDATORS -0.24897    0.46286  -0.538   0.5907
RAINLOW:TRATPREDATORS     0.04942    0.36944   0.134   0.8936
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

Correlation of Fixed Effects:
              (Intr) RAINLOW TRATNO TRATPR RAINLOW:TRATN
RAINLOW       -0.617
TRATNO_PRED   -0.703  0.440
TRATPREDATO   -0.817  0.510   0.581
RAINLOW:TRATN  0.410 -0.665  -0.584 -0.339
RAINLOW:TRATP  0.519 -0.833  -0.367 -0.632  0.554

library(car)
Anova(TELA_MANIP1,type="III",test="Chisq")

Analysis of Deviance Table (Type III Wald chisquare tests)

Response: ABUNDANCE
              Chisq Df Pr(>Chisq)
(Intercept) 61.8859  1  3.639e-15 ***
RAIN         0.0015  1   0.968899
TRAT        13.6932  2   0.001063 **
RAIN:TRAT    0.5580  2   0.756552
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

TELA_MANIP1<-glmer(ABUNDANCE~RAIN*TRAT+(1|SITE)+(1|caso),data=tela2013,poisson,glmerControl(optimizer
= c("Nelder_Mead")))

summary(TELA_MANIP1)

Generalized linear mixed model fit by maximum likelihood (Laplace
Approximation) ['glmerMod']
 Family: poisson  ( log )
Formula: ABUNDANCE ~ RAIN * TRAT + (1 | SITE) + (1 | caso)
   Data: tela2013
Control: glmerControl(optimizer = c("Nelder_Mead"))

     AIC      BIC   logLik deviance df.resid
   265.8    282.2   -124.9    249.8       49

Scaled residuals:
     Min       1Q   Median       3Q      Max
-1.67425 -0.55993  0.04442  0.63968  1.61151

Random effects:
 Groups Name        Variance  Std.Dev.
 caso   (Intercept) 8.144e-02 0.285378
 SITE   (Intercept) 6.202e-05 0.007875
Number of obs: 57, groups:  caso, 57; SITE, 7

Fixed effects:
                          Estimate Std. Error z value Pr(>|z|)
(Intercept)               1.512854   0.004555   332.1  < 2e-16 ***
RAINLOW                   0.016625   0.004551     3.7 0.000259 ***
TRATNO_PREDATORS         -0.619082   0.004550  -136.1  < 2e-16 ***
TRATPREDATORS             0.243134   0.004552    53.4  < 2e-16 ***
RAINLOW:TRATNO_PREDATORS -0.254128   0.004549   -55.9  < 2e-16 ***
RAINLOW:TRATPREDATORS     0.042442   0.004551     9.3  < 2e-16 ***
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

Correlation of Fixed Effects:
              (Intr) RAINLOW TRATNO TRATPR RAINLOW:TRATN
RAINLOW       0.000
TRATNO_PRED   0.000  0.000
TRATPREDATO   0.000  0.000   0.001
RAINLOW:TRATN 0.000  0.000   0.000  0.000
RAINLOW:TRATP 0.000  0.000   0.001  0.000  0.000

library(car)
Anova(TELA_MANIP1,type="III",test="Chisq")

Analysis of Deviance Table (Type III Wald chisquare tests)

Response: ABUNDANCE
                 Chisq Df Pr(>Chisq)
(Intercept) 110319.413  1  < 2.2e-16 ***
RAIN            13.343  1  0.0002594 ***
TRAT         21372.929  2  < 2.2e-16 ***
RAIN:TRAT     3207.377  2  < 2.2e-16 ***
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

TELA_MANIP1<-glmer(ABUNDANCE~RAIN*TRAT+(1|SITE)+(1|caso),data=tela2013,poisson,glmerControl(optimizer
= c("Nelder_Mead")))

summary(TELA_MANIP1)

Generalized linear mixed model fit by maximum likelihood (Laplace
Approximation) ['glmerMod']
 Family: poisson  ( log )
Formula: ABUNDANCE ~ RAIN * TRAT + (1 | SITE) + (1 | caso)
   Data: tela2013
Control: glmerControl(optimizer = c("Nelder_Mead"))

     AIC      BIC   logLik deviance df.resid
   265.8    282.2   -124.9    249.8       49

Scaled residuals:
     Min       1Q   Median       3Q      Max
-1.67425 -0.55993  0.04442  0.63968  1.61151

Random effects:
 Groups Name        Variance  Std.Dev.
 caso   (Intercept) 8.144e-02 0.285378
 SITE   (Intercept) 6.202e-05 0.007875
Number of obs: 57, groups:  caso, 57; SITE, 7

Fixed effects:
                          Estimate Std. Error z value Pr(>|z|)
(Intercept)               1.512854   0.004555   332.1  < 2e-16 ***
RAINLOW                   0.016625   0.004551     3.7 0.000259 ***
TRATNO_PREDATORS         -0.619082   0.004550  -136.1  < 2e-16 ***
TRATPREDATORS             0.243134   0.004552    53.4  < 2e-16 ***
RAINLOW:TRATNO_PREDATORS -0.254128   0.004549   -55.9  < 2e-16 ***
RAINLOW:TRATPREDATORS     0.042442   0.004551     9.3  < 2e-16 ***
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

Correlation of Fixed Effects:
              (Intr) RAINLOW TRATNO TRATPR RAINLOW:TRATN
RAINLOW       0.000
TRATNO_PRED   0.000  0.000
TRATPREDATO   0.000  0.000   0.001
RAINLOW:TRATN 0.000  0.000   0.000  0.000
RAINLOW:TRATP 0.000  0.000   0.001  0.000  0.000

library(car)
Anova(TELA_MANIP1,type="III",test="Chisq")

Analysis of Deviance Table (Type III Wald chisquare tests)

Response: ABUNDANCE
                 Chisq Df Pr(>Chisq)
(Intercept) 110319.413  1  < 2.2e-16 ***
RAIN            13.343  1  0.0002594 ***
TRAT         21372.929  2  < 2.2e-16 ***
RAIN:TRAT     3207.377  2  < 2.2e-16 ***
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1


best wishes,
Jordi

  Jordi Moya-Lara?o
Functional and Evolutionary Ecology
Estaci?n Experimental de Zonas ?ridas - CSIC
Carretera de Sacramento s/n
La Ca?ada de San Urbano
04120-Almer?a
Spain

phone:+34 950281045 ext. 419
email: jordi at eeza.csic.es
www.eeza.csic.es/foodweb