Differences between get_predicted and get_predicted_ci for mixed models about insight HOT 3 OPEN

get_predicted_ci() is mostly (internally) used for those model classes that do not return any standard errors for predictions (i.e. when there's nothing like se.fit = TRUE). This also applies to merMod-models. For these models, we use something like:

M <- model.matrix(m1)
V <- vcov(m1)
se <- sqrt(diag(M %*% V %*% t(M)))

based on https://bbolker.github.io/mixedmodels-misc/glmmFAQ.html#lme4. Thus, when calling get_predicted_ci(), standard errors are always based on the above formula. get_predicted() returns standard errors from predict() where possible. That's why get_predicted.glmmTMB() relies on standard errors returned by predict.glmmTMB(), but when calling get_predicted_ci() - even for a glmmTMB model - the above formula is used. That explains the difference between standard errors for glmmTMB models returned by get_predicted() or get_predicted_ci().

I think this should be clearly documented, and that get_predicted() is preferred (we may even hide get_predicted_ci() from the users?) in order to get SE and CI.

from insight.

Tagging @bbolker just FYI.

from insight.

Not sure if the documentation I added is sufficient, so I'll keep this issue open.

from insight.