[R-sig-ME] Plotting residuals for GLMER model and zero counts
bbolker at gmail.com
Wed Nov 6 22:33:20 CET 2013
Francesco <fbromano at ...> writes:
> Dear R-ers,
> although a number of options exist out there to plot logit models, I
> can't seem
> to find one that works for glmer. What I would like to do is, in light of
> having found no interaction between the two fixed effects (which answers my
> research question), look at a plot to tell how well the model fits the data.
> The package LMERconveniencefunctions ver 2.0
> has a nice plot for lmer but this won't work on my model because R
> asks me to use glmer. It has a factor DV (Correct), two fixed factor IVs
> (Group and Syntax),
> and two random effect (ID and item).
> Here's the output of calling:
> model<- glmer(Correct ~ Group * Syntax + (Syntax + 1 | ID) + (Group + 1
> | item), data=...., family=binomial)
I was going to ask you to please generate a (small) reproducible example
(e.g. see http://tinyurl.com/reproducible-000 ), but I couldn't resist
the urge to show off a new feature of the development branch of lme4,
which is that you can use simulate() with a regular (g)lmer formula
to simulate data corresponding to a given mixed model ...
## generate factorial combinations
d <- expand.grid(Group=c("ns","nns"),Syntax=c("'s","of"),
## subsample randomly down to actual data size
d <- d[sample(nrow(d),size=451),]
## check ...
## parameters (approximated from your output)
params <- list(fixef=c(-9,-2,5,-1.5),theta=c(7,-1,7,3,-0.5,6))
form <- Correct ~ Group * Syntax + (Syntax + 1 | ID) + (Group + 1| item)
s <- simulate(form[-2], ## use one-sided formula
## turn response into a factor
d$Correct <- factor(s[],labels=c("InC","Cor"))
fit1 <- glmer(Correct ~ Group * Syntax +
(Syntax + 1 | ID) + (Group + 1| item),
dAug <- data.frame(d,res=residuals(fit1,"pearson"))
plot(fit1) ## the default plot is not too useful
## how bad does the fit get if we leave out the interaction?
fit2 <- update(fit1,.~.-Group:Syntax)
## you can also separate by item etc. pretty easily ...
> On a side note, how do I cope with cases where participants produced
> 0 counts to a cell? I understand Chi-square does not cope well with
> these. Should I leave them as is, or turn these into NAs, or
> eliminate as outliers?
Leave them alone. You may be misunderstanding the framework of
GLMMs, which does *not* require a balanced design (although some
'structural' kinds of missingness, such as missing combinations
of interactions of *fixed* factors, will cause trouble) -- and
it's not based on a decomposition of a Pearson chi-squared table,
if that's what you're thinking ...
More information about the R-sig-mixed-models