[R-sig-ME] Passing lme4 optimizer arguments to pbmodcomp and afex/mixed
Tom Wenseleers
Tom.Wenseleers at bio.kuleuven.be
Wed Sep 25 17:48:57 CEST 2013
Hi Ben & Henrik,
Many thanks for the feedback!
(@Henrik: using your lme4 version with 10000 pwrssUpdate steps also didn't help unfortunately)
In case you would have the time to look into it, here is my example:
#library(pbkrtest) # latest dev version from http://people.math.aau.dk/~sorenh/software/pbkrtest/devel/
library(devtools)
install_github("lme4",user="lme4")
install.packages("afex",repos="http://r-forge.r-project.org",type="source")
library(afex)
data=read.csv("http://www.kuleuven.be/bio/ento/temp/alates.csv",header=T,fill=T)
summary(data)
fit=glmer(ALATES~ANT_TENDED*MELEZITOSE+NR_START+(1|CLONES)+(1|PLANT)+(1|REPLICATE),family=poisson,data=data)
summary(fit) # this works OK
# Generalized linear mixed model fit by maximum likelihood ['glmerMod']
# Family: poisson ( log )
# Formula: ALATES ~ ANT_TENDED * MELEZITOSE + NR_START + (1 | CLONES) + (1 | PLANT) + (1 | REPLICATE)
# Data: data_alates
#
# AIC BIC logLik deviance
# 741.1707 761.6854 -362.5853 725.1707
#
# Random effects:
# Groups Name Variance Std.Dev.
# PLANT (Intercept) 0.7423 0.8616
# REPLICATE (Intercept) 0.2388 0.4886
# CLONES (Intercept) 0.6739 0.8209
# Number of obs: 96, groups: PLANT, 96; REPLICATE, 12; CLONES, 8
#
# Fixed effects:
# Estimate Std. Error z value Pr(>|z|)
# (Intercept) 0.69149 1.61200 0.429 0.668
# ANT_TENDED1 0.15985 0.09792 1.632 0.103
# MELEZITOSE1 -0.28012 0.30861 -0.908 0.364
# NR_START 0.15771 0.16280 0.969 0.333
# ANT_TENDED1:MELEZITOSE1 0.08052 0.09706 0.830 0.407
#
# Correlation of Fixed Effects:
# (Intr) ANT_TENDED1 MELEZI NR_STA
# ANT_TENDED1 -0.134
# MELEZITOSE1 0.057 -0.010
# NR_START -0.978 0.136 -0.060
# ANT_TENDED1: 0.028 0.023 -0.005 -0.029
# LRTs also work:
mixed(ALATES~ANT_TENDED*MELEZITOSE+NR_START+(1|CLONES)+(1|PLANT)+(1|REPLICATE),family=poisson,type=3,method="LRT",data=data_alates)
#Fitting 6 (g)lmer() models:[......]
# Effect df.large df.small chisq df p
#1 (Intercept) 8 7 0.18 1 0.6744
#2 ANT_TENDED 8 7 2.59 1 0.1074
#3 MELEZITOSE 8 7 0.78 1 0.3783
#4 NR_START 8 7 0.89 1 0.3444
#5 ANT_TENDED:MELEZITOSE 8 7 0.68 1 0.4109
But if I try parametric bootstrapping it throws convergence errors:
cl <- makeCluster(rep("localhost", 4))
mixed(ALATES~ANT_TENDED*MELEZITOSE+NR_START+(1|CLONES)+(1|PLANT)+(1|REPLICATE),family=poisson,method="PB",args.test = list(nsim = 10,cl=cl),data=data_alates)
# Error in checkForRemoteErrors(lapply(cl, recvResult)) :
# 4 nodes produced errors; first error: pwrssUpdate did not converge in 10000 iterations
Same if one tries this from within lme4:
mm1=glmer(ALATES~ANT_TENDED*MELEZITOSE+NR_START+(1|CLONES)+(1|PLANT)+(1|REPLICATE),family=poisson,data=data)
mm <- model.matrix(mm1)
mm1 <- glmer(ALATES~ANT_TENDED*MELEZITOSE+NR_START+(1|CLONES)+(1|PLANT)+(1|REPLICATE),family=poisson,data=data)
form <- reformulate(c("-1",colnames(mm)[-1],"(1|CLONES)","(1|PLANT)","(1|REPLICATE)"),response="ALATES")
newdat <- data.frame(list(model.matrix(mm1)[,-1],getME(mm1,"flist"),model.frame(mm1)["ALATES"]),check.names=FALSE)
mm2 <- update(mm1,formula=form,data=newdat)
confint(mm2,method="boot",nsim=20)
#Computing bootstrap confidence intervals ...
# 2.5 % 97.5 %
#sd_(Intercept)|PLANT 0.69800722 1.0135147
#sd_(Intercept)|REPLICATE 0.19014163 0.6788588
#sd_(Intercept)|CLONES 0.56855996 1.0145768
#ANT_TENDED1 0.01637193 0.3186532
#MELEZITOSE1 -0.81867836 0.1685135
#NR_START 0.13921792 0.2738027
#ANT_TENDED1:MELEZITOSE1 -0.02224318 0.1300941
There were 27 warnings (use warnings() to see them)
> warnings()
Warning messages:
1: In optwrap(object at optinfo$optimizer, ff, x0, lower = lower, ... :
convergence code 1 from bobyqa: bobyqa -- maximum number of function evaluations exceeded
2: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
3: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
4: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
5: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
6: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
7: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
8: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
9: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
10: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
11: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
12: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
13: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
14: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
15: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
16: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
17: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
18: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
19: In pwrssUpdate(pp, resp, tolPwrss, GQmat, compDev, fac, ... :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
20: In bootMer(object, bootFun, nsim = nsim, ...) :
some bootstrap runs failed (10/20)
21: In norm.inter(t, alpha) : extreme order statistics used as endpoints
22: In norm.inter(t, alpha) : extreme order statistics used as endpoints
23: In norm.inter(t, alpha) : extreme order statistics used as endpoints
24: In norm.inter(t, alpha) : extreme order statistics used as endpoints
25: In norm.inter(t, alpha) : extreme order statistics used as endpoints
26: In norm.inter(t, alpha) : extreme order statistics used as endpoints
27: In norm.inter(t, alpha) : extreme order statistics used as endpoints
I presume it has to do with some simulated datasets not allowing parameters to be accurately estimated. I wonder if it wouldn't be possible perhaps to let PBmodcomp just ignore the ones for which no convergence could be achieved (or else to conservatively assume that the simulated LRT value in those cases was more extreme than the observed ones - which is probably better and less likely to lead to strong biases - at least you wouldn't have strong evidence that in those cases the LRT value for the simulated dataset was smaller than the observed one), and throw a warning in which it would be reported in how many cases no convergence could be achieved (if this number of small, the result might still be more or less usable). Or is there something that could be done on the lme4 side with respect to optimizer settings to get over this? Also, is there perhaps some provisions to calculate p-values for all factors from the lme4 likelihood profiles?
Cheers,
Tom
-----Original Message-----
From: Henrik Singmann [mailto:henrik.singmann at psychologie.uni-freiburg.de]
Sent: 25 September 2013 15:36
To: Ben Bolker
Cc: Tom Wenseleers; Søren Højsgaard; r-sig-mixed-models at r-project.org
Subject: Re: Passing lme4 optimizer arguments to pbmodcomp and afex/mixed
I am also currently working on a binomial GLMM on which I encounter this problem regularly (depending on which model I fit). But as my data is somewhat large (12000 observations) and fitting takes a while I wouldn't consider it a nice demonstration of the issue.
However, I did try to increase the number of steps in pwrssUpdate by altering the source of lme4 and, as predicted by Ben, this does not help at all (it just reaches the new boundary).
But, if nevertheless someone is interested to try it out for him or herself, you can install a version of lme4 1.1.0 that is identical to the development version as of yesterday that only stops after 10000 pwrssUpdate steps from my github:
library("devtools"); install_github("lme4",user="singmann")
Best,
Henrik
-----Original Message-----
From: Ben Bolker [mailto:bbolker at gmail.com]
Sent: 25 September 2013 15:21
To: Tom Wenseleers; Søren Højsgaard; henrik.singmann at psychologie.uni-freiburg.de; r-sig-mixed-models at r-project.org
Subject: Re: Passing lme4 optimizer arguments to pbmodcomp and afex/mixed
At the moment it's not possible to increase the number of iterations; that's on the list to allow, but my guess is that that's not going to help. Unfortunately, our test cases for lme4 did not include a lot of binomial GLMMs with complete separation and/or extreme values of the beta parameters (say |beta|>5). I suspect that what's going on here is similar to the examples listing in these issues ..
https://github.com/lme4/lme4/issues/134
https://github.com/lme4/lme4/issues/124
Is it in fact the case that if you fit this model with lme4.0 (or an older version of lme4) you get large values of beta?
Looks like we're going to have to put higher priority on fixing this ...
Ben Bolker
On 13-09-25 09:13 AM, Tom Wenseleers wrote:
> Dear all, Just a quick question re pbmodcomp (I'm using version
> 0.4-0 of pbkrtest, and latest devel version of lme4). I am calling
> pbmodcomp via afex, but I am getting the dreaded error "3 nodes
> produced errors; first error: pwrssUpdate did not converge in 30
> iterations" (my call is
> mixed(ALATES~ANT_TENDED*MELEZITOSE+NR_START+(1|CLONES)+(1|PLANT)+(1|RE
> PLICATE),type=3,family=poisson,method="PB",args.test
> = list(nsim = 10,cl=cl),data=data_alates) ).
>
> I would now like to try passing some optimizer arguments to the lmer
> calls. To do this I first called options(glmerControl=list(optimizer
> = c("bobyqa", "bobyqa"),tolPwrss = 1e-03
> ,optCtrl=list(maxfun=1000))) But then I still get " pwrssUpdate did
> not converge in 30 iterations". Does one of you happen to know how I
> could perhaps change the nr of maximum iterations to >30? And also if
> it is possible to pass such arguments in the call to mixed and/or
> pbmodcomp?
>
> Cheers, Tom
>
>
> ______________________________________________________________________
> _________________
>
> Prof. Tom Wenseleers P Lab. of Socioecology and Social
> Evolution Dept. of Biology Zoological Institute K.U.Leuven
> Naamsestraat 59, box 2466 B-3000 Leuven Belgium +32 (0)16 32 39 64 /
> +32 (0)472 40 45 96 //tom.wenseleers at bio.kuleuven.be
> http://bio.kuleuven.be/ento/wenseleers/twenseleers.htm
>
>
>
More information about the R-sig-mixed-models
mailing list