[R-sig-ME] Fw: lme4

[Ben Bolker]

Ebi Safaie <safaie124 at ...> writes:

> Dear Ben Bolker, 
> Thank you very much for your informative reply.
> Yes, I followed Barr et al (2013).
> 
> I did what you kindly sent me. I'm not sure I've
> done it correctly but it came to false
> 
> It would be a good idea to check for a singular fit, i.e.
> 
>   t <- getME(mod.15,"theta")
>   lwr <- getME(mod.15,"lower")
>   any(t[lwr==0]< 1e-6)
> 
> t <- getME(mod.15,"theta") > lwr <- getME(mod.15,"lower") 
> any(t[lwr==0]< 1e-6) [1] FALSE

  that's good -- that means that your model is at least
bounded away from zero for constrained parameters.
 
> I increased the number of iterations as you suggested
> 
> summary(mod.15<-glmer(ErrorRate~1
> cgroup*cgrammaticality*cHeadNoun*cVerbType+(1|itemF)+
> (1+grammaticality*HeadNoun*VerbType|participantF),data=e3, 
> family="binomial",na.action=na.exclude,
> control=glmerControl(optCtrl=list(maxfun=1e6))))
> 
> but it came to the following message
> 
> Warning messages: 1: In checkConv(attr(opt, "derivs"), 
> opt$par, ctrl = control$checkConv,  : Model
> failed to converge with max|grad| = 0.113924 
> (tol = 0.001, component 29) 2: In checkConv(attr(opt,
> "derivs"), opt$par, ctrl = control$checkConv,  :
> Model failed to converge: degenerate  Hessian with 1
> negative eigenvalues   

  These warnings do suggest that your model is at the very least
unstably fitted. You could try some of the strategies listed
at 

http://rpubs.com/bbolker/lme4trouble1

to reassure yourself that the model fit is in fact OK.

I want to emphasize again that your model is **not** actually
fitting worse than it did before/with previous versions; rather,
the default warning level has been turned up so that you're
getting more warnings than before.

> Actually the following two interactions are important for me
>  because they are representing two hypothesis
> 2 way 
> 
> cgroup*cgrammaticality
> 
> 4 way interaction
> cgroup*cgrammaticality*cHeadNoun*cVerbType


Comparing previous results just for these terms --

previous
                                 est     stderr       Z        P
cgroup:cgrammaticality        1.5796     0.3586   4.404 1.06e-05 *** 
cgroup:cgrammaticality:       3.1326     1.3994   2.239   0.0252 *
  cHeadNoun:cVerbType

current

cgroup:cgrammaticality        1.57010    0.36695  4.279 1.88e-05 ***
cgroup:cgrammaticality:       3.15344    1.42351  2.215   0.0267 *
   cHeadNoun:cVerbType

As I said before, the new and old results
look the same to me for all practical
purposes.

> Earlier, I used odds ratio to calculate the effect sizes
> (Table below) and I was able to
> dissociate between these two interactions (i.e., two hypotheses) 
> via their effect sizes. 
> Due to wider range of the lower and upper limits of 95% CI 
> I supported the 2 way interaction. 

   Don't know what you mean here.  Are you trying to distinguish
which one has a larger effect?  Assuming all your predictors
are categorical (so that you don't have to worry about standardizing
units), the two-way interaction has a smaller _effect_ but also
smaller uncertainty, so it is more statistically significant.

> Am I on the right track? 
> Given that I want to use the newer version of lme4 (as you recommended) 
> I would really appreciate your help to let me know what to do 
>  with this really 
> complex design.
>   
>   Table 9.Experiment 1a: Fixed-effects from mixed-effects logistic 
> regression model fit to data from both
> NSs and the NNSs for S-V agreement  Main analysis 
> Fixed effects:           Odds Ratio (OR) 95% CI
> For OR 

Your table got somewhat mangled in transition to the mailing list,
but appears to be a slightly modified version of the summary() output,
with odds ratios and Wald confidence intervals on odds ratios (i.e.
based on exp(est +/- 1.96*std. err) appended).

   The questions about warning messages from lme4 and what to do
about them are on-topic for this list, but these questions about how to
interpret the fixed effects are pretty generic (e.g. they would
apply pretty much equivalently to a regular linear or generalized linear
model), and would be more appropriate for a more generic stats questions
venue such as CrossValidated <http://stats.stackexchange.com>

  sincerely
    Ben Bolker