[R-sig-ME] random slopes in gamm4

[Ben Bolker]

 See also:

Pedersen, Eric J., David L. Miller, Gavin L. Simpson, and Noam Ross.
“Hierarchical Generalized Additive Models: An Introduction with Mgcv.”
PeerJ Inc., November 5, 2018.
https://doi.org/10.7287/peerj.preprints.27320v1.


On 2019-04-06 4:37 a.m., Cesko Voeten wrote:
> Dear John,
> 
> A random slope in gamm4 is the same as a random slope in lme4, i.e.
> (X1|ID) gives you separate linear slopes for X1 by IDs. If X1 is
> numeric, this should be exactly identical to mgcv's s(ID,X1,bs='re').
> 
> It sounds like what you really want is a random smooth, which in mgcv is
> called a factor smooth: s(ID,X1,bs='fs',xt=list(k=4),m=1). If you want
> the *smooth* to vary across IDs, this is appropriate. Pass 'm=1' to
> penalize the smooth a little bit more -- this is common for random smooths.
> 
> You can read more about this on
> http://www.sfs.uni-tuebingen.de/~jvanrij/Tutorial/GAMM.html. If you use
> a factor smooth, you do not need to add a separate random slope, btw.
> 
> Best,
> Cesko
> 
> Op 05-04-2019 om 21:44 schreef René:
>> Dear John,
>>
>> I almost tend to say, you should ask Simon Wood, Fabian Scheipl
>> directly :))
>> but, you know, in the gamm4 manual for formula they say: 'Note that
>> ids for
>> smooths and fixed smoothing parameters are not supported.'
>> (Which I do not understand) But later it says:
>>
>> gamm4 allows the random effects specifiable with lmer to be combined with
>> any number of any of the (single penalty) smooth terms available in
>> gam from
>> package mgcv as well as t2 tensor product smooths. Note that the model
>> comparison on the basis of the (Laplace approximate) log likelihood is
>> possible with GAMMs fitted by gamm4.
>>
>> Which I, honestly, again, absolutely do not understand. But it seems that
>> random slopes are about penalty smooth terms (how much you like them
>> to...). Interestingly, when clicking on 'gam' (to learn more), it
>> redirects
>> me to the same page... which almost feels like magic (or buddhist).
>>
>> And the second part of your question alarms me...  it seems that you want
>> to analyze residual terms (i.e. random errors, which by frequentist
>> definition are random), is this the case? Otherwise, whether positive or
>> negative, the mean of any random effect should be zero, and the residuals
>> should be normally (or appropriately) distributed. I would argue (as
>> others
>> do), that any directed hypothesis should be part of the fixed effects.
>> And
>> maybe the question is, whether there are better packages to do this, and
>> allow to predict the function parameters you are interested in (In
>> case of
>> interest, check 'brms' and its non-linear function applications. You
>> might
>> become happier with this, as it would be on you to define 'positive or
>> negative', and also with respect to the support.)
>>
>> :)
>>
>>
>> Best wishes, René
>>
>>
>>
>> Am Fr., 5. Apr. 2019 um 14:48 Uhr schrieb John Morrongiello <
>> john.morrongiello using unimelb.edu.au>:
>>
>>> Hi all
>>>
>>>
>>>
>>> I posted this question back in 2015 but unfortunately didn't get a
>>> reply.
>>> I'd like to use these models again so thought it worth another ask.
>>>
>>>
>>>
>>> Most GAMM examples involve the fitting of a model with just a random
>>> intercept (e.g. M1 below). However, I'd like to explore the
>>> possibility of
>>> each individual (ID) having a different X1 smoother 'slope', or even
>>> just
>>> linear slope, akin to a random slope in lmer (M2).
>>>
>>>
>>>
>>> M1<-gamm4(response ~ s(X1,k=4), random =~(1|ID),data)
>>>
>>> M2<-gamm4(response ~ s(X1,k=4), random =~(X1|ID),data)
>>>
>>>
>>>
>>> However, I'm unsure how to interpret the random slope X1 in M2. If
>>> positive, is it an overall increase in smoother 'wriggliness' i.e.
>>> increase
>>> in edf (opposite for negative)? Or is it something else? Would
>>> someone know
>>> how to visualise these 'random slopes' so I can get a feel for what is
>>> going on?
>>>
>>>
>>>
>>> A manuscript by Pedersen and colleagues has recently been posted on PeeJ
>>> that explores related models within the gam function
>>> https://peerj.com/preprints/27320/. I'm happy to use these, but still
>>> would like to understand the random slope in a gamm4 context.
>>>
>>>
>>>
>>> Cheers
>>>
>>>
>>>
>>> John
>>>
>>>          [[alternative HTML version deleted]]
>>>
>>> _______________________________________________
>>> R-sig-mixed-models using r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>
>>
>>     [[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