This is promising, I very frequently end up writing one off functions to munge data on CIs between lme4 and ggplot2.
However, what a prediction interval actually means is a slightly conentuous issue in the context of mixed models, and Douglas Bates had plenty of caution about constructing/interpreting them on the R mailing lists a few years back.
This said, I’ve always found the tip-toeing around constructing them annoying; it just means that there’s little standard practice.
Do you have any documentation on how you’re doing prediction intervals for mixed models, and what interpretations do and don’t make sense for your method?
Good question! There is a vignette in the GitHub version that specifically addresses mixed models and lists exactly which variances and standard errors are incorporated in interval estimation. We show in the vignette that our intervals reach the nominal coverage level asymptotically.
I'm confused by the use of ggplot2 as shown in the example: ggplot2 and geom_smooth() already calculates and plots confidence intervals by default. What does this package add to a ggplot?
ciTools doesn't add anything to ggplot, it creates tidy dataframes that are easy to plot with ggplot. In the example, we show CIs and PIs on a linear mixed model, which ggplot can't handle.
ggplot's has a built in interval for lm that I'm aware of, but there are essentially statistical interval toolkits that you can supply yourself. This lib is a handful of implementations.
Bayesian stats has "credible intrrvals." These are different than those from relative likelihood and those from Frequentist methods.
ggplot's built-in scratch the surface but are still really handy and easy to extend.
However, what a prediction interval actually means is a slightly conentuous issue in the context of mixed models, and Douglas Bates had plenty of caution about constructing/interpreting them on the R mailing lists a few years back.
This said, I’ve always found the tip-toeing around constructing them annoying; it just means that there’s little standard practice.
Do you have any documentation on how you’re doing prediction intervals for mixed models, and what interpretations do and don’t make sense for your method?