[R-sig-ME] same model runs in nlme but not lme4

[Simon Harmel]

Many thanks!

On Fri, May 22, 2020 at 6:35 PM Ben Bolker <bbolker using gmail.com> wrote:

>    They're pretty separate things.  The  likelihood profile is completely
> conditional on the model.  I suppose if the data are completely insane then
> the profile will probably be weird too.  The profile has to do with the
> shape of the likelihood surface rather than the distribution of the
> variation around the model.
> On 5/22/20 7:24 PM, Simon Harmel wrote:
>
> Short but very clear. Appreciate it very much. Don't mean to make this
> long, but how this likelihood profile analysis relates with fitted vs.
> residual relation? Can they be at odds?
>
> On Fri, May 22, 2020 at 5:41 PM Ben Bolker <bbolker using gmail.com> wrote:
>
>>
>>    Profile plots expressed in terms of the signed square root are
>> straight lines if the log-likelihood surface is quadratic (in which case
>> the Wald confidence intervals will be reliable). (I know that's very terse
>> but I'm composing in haste.)
>>
>>   vignette("lmer", package="lme4") has a little bit.  More generally you
>> can read in any advanced stats book about likelihood profiles and what they
>> are/mean (section 4 of https://ms.mcmaster.ca/~bolker/emdbook/chap6A.pdf
>> gives one such introduction).
>> On 5/22/20 6:35 PM, Simon Harmel wrote:
>>
>> Many thanks, Ben. Just curious, what information do the plots at the end
>> of your exactly convey?
>>
>> I also appreciate it if there if you could point me to a documentation in
>> lme4 where I can learn more about `profile()` and its output.
>>
>> Many thanks, Simon
>>
>> On Fri, May 22, 2020 at 5:25 PM Ben Bolker <bbolker using gmail.com> wrote:
>>
>>>    Because lme4 is fussier than lme.  lme will fit models where the
>>> variance components are jointly unidentifiable; lmer tries to detect
>>> these problems and complains about them.  It's possible that this is a
>>> false positive.  You can make it run by specifying
>>>
>>> m1 <- lmer(y~ group*year + (year|stid), data = dat,
>>> control=lmerControl(check.nobs.vs.nRE="ignore"))
>>>
>>>    but I strongly recommend that you think about whether this might be
>>> exposing problems.
>>>
>>>   calculating the profile suggests a little bit of weirdness.
>>>
>>> pp <- profile(m1,signames=FALSE)
>>>
>>> dd <- as.data.frame(pp)
>>>
>>> library(ggplot2)
>>> ggplot(dd,aes(.focal,.zeta)) + geom_point() + geom_line() +
>>> facet_wrap(~.par,scale="free_x")
>>>
>>> You can compare confint(pp) to intervals(m2); they're mostly consistent,
>>> but some caution is suggested for the CIs on the correlation and the
>>> year SD
>>>
>>>
>>> On 5/22/20 5:57 PM, Simon Harmel wrote:
>>> > Hi All,
>>> >
>>> > I was wondering why my model runs ok when I use `nlme` package but it
>>> fails
>>> > when I use the `lme4` package, am I missing something?
>>> >
>>> > Thanks, Simon
>>> >
>>> > #===================================
>>> > library(lme4)
>>> > library(nlme)
>>> >
>>> > dat <- read.csv('https://raw.githubusercontent.com/hkil/m/master/z.csv
>>> ')
>>> >
>>> > m1 <- lmer(y~ group*year + (year|stid), data = dat)       ## Fails ###
>>> >
>>> > m2 <- lme(y~ group*year, random = ~year|stid, data = dat) ## Runs ###
>>> >
>>> >       [[alternative HTML version deleted]]
>>> >
>>> > _______________________________________________
>>> > R-sig-mixed-models using r-project.org mailing list
>>> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>
>>> _______________________________________________
>>> R-sig-mixed-models using r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>
>>

	[[alternative HTML version deleted]]