what do plevel and showCI do for sjp.int with mixed models? about sjplot HOT 8 CLOSED

plevel only applies to models that do actually compute p-values. I think, glmer does produce p-values. In sjp.lmer, I use car::Anova to get approximate p-values, however, I cannot apply this function properly to sjp.int because I don't know how to deal with factors - the anova computes a p-value for the whole predictor, not for each factor level (thus, you would have to create dummy values for each factor level).

showCI only works for plot types eff and emm in sjp.int, just because the packages I use to compute these interactions provide CI-values. I don't know how to compute confidence intervals on conditional interaction / moderation effects - I appreciate any help!

(see this sjPlot-manual for what I understood as "conditional effect")

from sjplot.

lmerTest and mixed do p-values for lmer (for each level), but of course Douglas Bates has some doubts about those (and I think when I use the function I prefer to just plot all defined interactions without regard to p-values (i.e. set plevel to 1), so I can compare different plots visually).

I didn't notice showCI works for eff, I don't have smart ideas about how to compute CIs for "cond" either. Maybe you could put this in the docs, though?

"emm" with lmer fails for me, if I use lmerTest I get

Error in summary(fit)$coefficients[-1, 4] : subscript out of bounds

and with plain lme4

Error in colnames<-(*tmp*, value = c("x", "y", "grp", "l.ci", "u.ci", :
'names' attribute [6] must be the same length as the vector [5]

So you pass the fits to another package to compute CIs?

Maybe you can put this on the wishlist then: I would sometimes like to customise options for predictions:

1. e.g. I prefer to use bootstrapping for SEs/CIs/predictions and
2. one use case for which I wrote my own prediction function for interactions once involved extrapolating (i.e. going over the range of the real data in newdata) and
3. in some cases choosing the mean of the covariates for "eff" is not unambiguous (e.g. three equally large birth cohorts) and the user might want a choice

If I get a handle on your codebase, maybe I can contribute something along those lines, I really like the package and dislike that I'm always writing my own not-very-reusable functions for plotting coefficients, predictions etc.

from sjplot.

@1 I have not much experience with bootstrapping yet, so I'm not sure what you are thinking of exactly, and how to implement it?

@2 Would be possible for type = "cond", but I'm not sure how to set custom ranges for predictors in the effects package / effect function (which is the base for type = "eff")?

@3 Which other options may be useful? You can specify the "averaging" effect via the typical parameter (in effect), which must be a function. Median?

from sjplot.

wrote you an email regarding 1.
Re2: I haven't used effects before, but it seems like this (and the bool problem) would be some things to submit to the maintainer?

Re3: I wasn't familiar with that, thanks!

from sjplot.

mixed models and type = "emm" should work now. CI's are calculated by the lsmeans package.

from sjplot.

What about this issue? What parts are still "open"?

from sjplot.

This works now. You could get pvalues from lmerTest instead of anova.

from sjplot.

Only if you fit the model with lmerTest::lmer, afaik. There's no method in the lmerTest package to obtain p-values from an merMod-object from lame. So, you have to use lmerTest to fit your models, and if you do so, sjPlot will take those p-values provided by lmerTest.

from sjplot.