[R-sig-ME] Fwd: R: random structure in lmer

Ben Bolker bbolker at gmail.com
Tue Feb 11 16:07:46 CET 2014


  cc'ing back to r-sig-mixed-models ...

  The example you showed in your original post did not have random
slopes, which might explain my confusion ...

  Next question: when you fit the random-slopes model, do you have a
singular fit (i.e. do you have variances estimated to be zero, or
correlations estimated to be +/- 1.0, or are some of the elements of
getME(fitted_model,"theta") equal to zero)?

  I think this would explain a non-positive-definite Hessian matrix
(i.e., matrix of second derivatives of the fitted parameters).

  If that is the case, there is a lot of discussion on this list about
what to do in this situation, although no entirely satisfactory answers.
(There's not much very specific at http://glmm.wikidot.com/faq ; I
should add a section.)

  Ben Bolker


-------- Original Message --------
Subject: R: [R-sig-ME] random structure in lmer
Date: Tue, 11 Feb 2014 16:01:22 +0100
From: Cerni, Tania <Tania.Cerni at unitn.it>
To: Ben Bolker <bbolker at gmail.com>

Thanks for the answer,
Well, I receive this message every time I introduce a random slope in
the model. Not with the simple structure  (1|Sj)+ (1|Word).
In the experiment I'm interested in this interaction:
w.pw*group*manipulation or group*manipulation. The other variables are
inside only to see if they have an effect that can potentiate or delete
the interaction (for example "letters" is the length of the word,
"Trial" is the trail number to see an effect of fatigue) and I do not
think about multicollinearity.
Do you think I have to investigate only the interaction without
controlling words parameters?

Tania

-----Messaggio originale-----
Da: Ben Bolker [mailto:bbolker at gmail.com]
Inviato: martedì 11 febbraio 2014 14:53
A: Cerni, Tania; r-sig-mixed-models at r-project.org
Oggetto: Re: [R-sig-ME] random structure in lmer

On 14-02-11 06:07 AM, Cerni, Tania wrote:
> Dear all, I'm  a PhD student in Cognitive science. I'm working on an 
> analysis with mixed models with R (lme4 and lmerTest)  but I'm  still 
> in trouble with random factors. I have read some Baayen's works and I 
> tried to follow his methods, but every time, after comparing models 
> with different slopes, the output of lmer is the summary of the model 
> preceded by:
> 
> "Asymptotic covariance matrix A is not positive!"
> 
> I want to ask you what kind of problems can I have with this message 
> and, most of all, I'm writing to you to ask if you can advise me a 
> shared procedure for random structure.

  For what it's worth, this message comes from lmerTest, not lme4, and I
suspect it actually has to do with your fixed-effect structure, or the
interaction between your fixed-effect structure and your random-effect
structure, not your random-effect structure per se.  I suspect there is
some near-perfect multicollinearity among some of your factors, but I'm
not sure.  Have you thought carefully about your experimental design to
make sure that all of the parameters you want to estimate are really
identifiable ... ?

   Ben Bolker

> 
> I have this kind of models: with 3way interaction or 2way interaction.
> 
> lex0 = lmer(rt~Trial +first_letter+letters+w.pw*group*manipulation+
> (1|Sj)+ (1|Word),lex)
> 
> the main thing is to see if there is an interaction between groups and 
> manipulation.  The other fixed factors are predictors
> 
> 
> 
> Random effects:
> 
> Groups   Name        Variance Std.Dev.
> 
> Word     (Intercept) 0.006321 0.0795
> Sj       (Intercept) 0.022424 0.1497
> Residual             0.032240 0.1796
> 
> Number of obs: 7169, groups: Word, 240; Sj, 32
> 
> 
> 
  [snip]


> How can I adjust my random structure? Thanks, Tania



More information about the R-sig-mixed-models mailing list