Date: 2012-11-17 15:04
Sender: Stéphane Laurent
Here is an example where lme4 development version gives almost the same results than SAS whereas the stable version gives slightly different results. The csv file is available in the attachments.
dd <- read.csv("datlme4.csv")
dd <- transform(dd, sample=factor(sample))
# lme4 cran :
lmer(log10(y) ~ (1|sample)+(1|operator)+(1|sample:operator), data=dd)
# lme4 development :
lmer(log10(y) ~ (1|sample)+(1|operator)+(1|sample:operator), data=dd,
control=list(restart=TRUE), optimizer="bobyqa")
|
Date: 2012-07-08 19:39
Sender: Ben BolkerI've looked a bit at the second example and can't see that it has anything to do with boundary issues, rather it seems just to be a bobyqa/Nelder-Mead contrast. This is really tough. We have Nelder-Mead as a default: changing the default generically to bobyqa is a big deal and would require very careful thought and testing ...
|
Date: 2012-07-05 08:33
Sender: Stéphane LaurentI have an example where BOTH arguments optimizer="bobyqa" and control=list(restart=TRUE) are required in order to get the good results.
I have attached the dataset, and the model is :
dat2 <- read.csv("dat2009_0551.csv")
lmer(y ~ (1|Sample)+(1|Day)+(1|Operator)+(1|Day:Sample)
+(1|Day:Operator)+(1|Sample:Operator)+(1|Day:Sample:Operator), data=dat2) |
Date: 2012-07-04 06:35
Sender: Stéphane LaurentI confirm that this works fine with either optimizer="bobyqa" or control=list(restart=TRUE).
Thanks ! |
Date: 2012-07-03 12:34
Sender: Ben BolkerI'm pretty sure this is another case where Nelder-Mead is getting stuck at the boundary by accident. If so, then using optimizer="bobyqa" should be a temporary workaround; so should control=list(restart=TRUE) in the latest development version. I will play around with some ideas for automatic restarting from the boundary (my initial attempt was too crude and led to some problems with existing tests ...)
|
