I am aware that with confint it is possible to extract a confidence intervall of the difference of the levels of the effect, but not a confidence intervall on a single level.
Do you mean the "Beta" values. There is a confint method for nerMod objects (what are created by glmer), so that would work. If you mean the variability in the conditional modes, use ranef(model,varCond=TRUE). For the manual see https://cran.r-project.org/web/packages/lme4/lme4.pdf or the Journal of Statistical Software paper. Another very good source is Doug Bates' draft book (since he is working on Julia it may not get finished) on github https://github.com/dmbates.
The easiest way to do it is to run the model with that level set to the reference level (and all categorical predictors dummy-coded). Then the intercept corresponds to the estimate for that level, and you can use confint or bootMer to get the CI of that estimate.
If you have just one predictor and you dummy-code it and run an interceptless model, then you will get estimates for each level of the predictor, rather than for differences between the levels. e.g.,
many thanks for your suggestions @Daniel, I did not mean the beta values @Stephen, in my case my predictors would have only two level, therefore I will not care about dummy coding.
I will give an example of my problem to make myself more clear:
I have a depending variable with a skewed distribution, let's call it "absolute frequency". Furthermore, I have a within-subject factor behavior with two levels (A and B) and a between-subject factor group (patients and healthy participants). I decided to compute a glmer with poisson distribution instead of a repeated measures ANOVA. My model looks like this:
Like I mentioned, if you dummy-code then the intercept refers to the reference level and you can easily get the confidence interval of it. (This is the case even when your factors have two levels; even with two levels, dummy coding still gives different coefficients than e.g. deviation coding, so I don't see why having two-level predictors means you "will not care about dummy coding".)
The easiest option is to recode your data into a 1x4 design rather than a 2x2 design. e.g., create a new variable called "Condition" with four levels:Apatient, Ahealthy, Bpatient, Bhealthy. Then run an interceptless model with that factor dummy-coded:
Thank you Stephen, the approach with the dummy coding worked very well.
Now, I would like to extract the accurate confidence intervals in a similar analysis where my dependent variable is not a absolute, but a relative frequency. Something like behavior in relation to the number of words produced by the participant? For this purpose, I used the a offset:
One of the best ways of getting CIs from glmer fits is parametric bootstrapping, although it can be slow. See: http://glmm.wikidot.com/faq
Off-topic, but you didn't show the random effects specification and I wondered if you had included a term to allow for overdispersion in the model? If the model needs such a term and it isn't included, then the CIs will tend to be over-optimistic (i.e. too narrow). An easy way to do this in glmer is to add a factor of the sort