[R-sig-ME] parsimonious mixed models

[Phillip Alday]

Having an interaction without the corresponding main effect is a bit unusual, but not unheard of ... in your case though, I would remove the interactions and take look at the new random-effects estimates. I suspect you'll see that more variance is now explained by the main effects. 

I also suspect that your variables aren't quite coded right -- cSenttype is probably categorical, is it not? Yet it's estimate in the random-effects structure is that of a continuous variable (single estimate) and not a factor (set of estimates for different levels / contrasts).

Best,
Phillip

> On 9 Nov 2016, at 21:14, Wing Yee Chow <wingyeechow.zoey at gmail.com> wrote:
> 
> Hi there, 
> 
> I'm trying to follow Bates et al.'s guidelines for parsimonious mixed models
> and have two questions:
> 
> 1.       In reducing (near-zero) variance components from the random effects
> structure, is the criterion simply the size of the variance? Or do I need to
> keep a simple effect as long as I'm keeping a higher order effect that
> involves that factor? That is, with the following random effects structure,
> do I take out (i) just cSenttype:cCongruity for item, or (ii) both
> cSenttype:cCongruity for item AND cCongruity for subj?
> 
> 
> 
> Random effects:
> 
> Groups   Name                 Variance Std.Dev.
> 
> item     cSenttype:cCongruity       0     0.0  
> 
> item.1   cCongruity            205437   453.3  
> 
> item.2   cSenttype             100765   317.4  
> 
> item.3   (Intercept)           124422   352.7  
> 
> subj     cSenttype:cCongruity  134752   367.1  
> 
> subj.1   cCongruity                 0     0.0  
> 
> subj.2   cSenttype              87161   295.2  
> 
> subj.3   (Intercept)          2434746  1560.4  
> 
> Residual                      4891977  2211.8  
> 
> Number of obs: 1082, groups:  item, 48; subj, 24
> 
> 
> 
> 2.       In another dataset, I have a 3 x 2 within-participant and
> within-item design. I specified the contrasts for the two factors in this
> way:
> 
> contrasts(tempdata$congruity) <- contr.sum(2)/2
> 
> contrasts(tempdata$sentencetype)=cbind("AvsBC" = c(-2/3, 1/3, 1/3), "BvsC" =
> c(0, -1/2, 1/2))
> 
> 
> 
> I got the model matrix (mmatrix) from the maximal model (m0) to construct
> the zero-correlation parameter model (m1). I think I know how to do this for
> a factor with two levels (i.e., only one contrast), but I'm not sure about
> the current case since I have both cSenttypeAvsBC and cSenttypeBvsC for the
> 3-level factor sentencetype. Is this the right syntax?
> 
> 
> 
> cSenttypeAvsBC <- mmatrix [,2] # columns 2 and 3 encode different contrasts
> of sentencetype
> 
> cSenttypeBvsC <- mmatrix[,3]
> 
> cCongruity <- mmatrix[,4]
> 
> m1<- lmer(value ~ sentencetype * congruity + ((cSenttypeAvsBC +
> cSenttypeBvsC) * cCongruity||subj) + ((cSenttypeAvsBC + cSenttypeBvsC) *
> cCongruity||item), REML=FALSE,
> 
>                          data=tempdata)
> 
> 
> 
> Many thanks!!
> 
> Wing-Yee Chow
> 
> 
> 	[[alternative HTML version deleted]]
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models