[R-sig-ME] glmer and lme4 - Quick question

[Ken Beath]

What 0+ is doing is removing the intercept random effect. It is then
included with the 1|Region term, however there is now no correlation
between the slope and intercept random effects. I can understand why you
don't want to do that as including something like (x1+x2+x3|Region) means
estimating a large covariance matrix for the random effects. However the
bad news if the correlations are not zero then it will give lots of
estimation problems and probably some strange results. Your convergence
problems may well be due to not having a complex enough model.

Opinions vary on what to do. Either start with something simple and keep
adding, or start with all the random effects and remove until it converges
and then remove the unimportant random effects.

On 24 June 2015 at 10:10, Joseph Maina <mainajm at gmail.com> wrote:

> Hi,
> I am running glmer in a model selection framework with >30 explanatory
> variables, where I am first  generating all possible combinations of
> variables but with a multicollinearity test results coinstraint,  before
> fitting the gmler. I am also including random effects of ‘Year' (~20) and
> Regions (~13). The objective is to find the best model and determine the
> relative influence among predictors, and also to predict the model over
> space to global pixels.
>
> My question regards the model structure of lme4 that I should adopt. I am
> currently fitting my model in the following structure:
> m1<-glmer(y  ~ x1 + x2 + x3 + (1 |Region) + (1 | Year) +(1|Region),
> family=binomial('logit'),data=all.data)
>
> However, I have been advised that in order to have a varying intercept and
> slope among my Regions (one of the random effects), I should fit my model
> as follows:
> m1<-glmer(y ~ x1+ (0+x1|Region) + x2 + (0+x2|Region) + x3 + (0+x3|Region)
> + (1 | Year) +(1|Region), family=binomial('logit'),data=all.data)
>
> The latter is a slightly complex structure and I am running into
> convergence issues. I was wondering what are the merits of using either
> structure?.  Also in the second structure, I am not sure what ‘0+’ means or
> what value it adds to the analyses. I also found that when using the first
> model structure if I take out the ‘Region’ random effect, the estimates for
> some of the variables change signs, and therefore could have an
> implicaition on the interpretation.
>
> Thanks,
>
> Joseph
>
>
>         [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>



-- 

*Ken Beath*
Lecturer
Statistics Department
MACQUARIE UNIVERSITY NSW 2109, Australia

Phone: +61 (0)2 9850 8516

Level 2, AHH
http://stat.mq.edu.au/our_staff/staff_-_alphabetical/staff/beath,_ken/

CRICOS Provider No 00002J
This message is intended for the addressee named and may...{{dropped:9}}