The parametric bootstrap (boot.type = "parametric")
for lmer and rlmer fits now generates both the
random effects and the errors anew in every bootstrap sample, as
described in Mason, Cantoni & Ghisletta (2021). In earlier versions
the random effects were held fixed at their predicted values and only
the errors were redrawn, which gave intervals that were too narrow, most
notably for the fixed effects and the random effects’ standard
deviations. This affects method = "boot" and
method = "BCa". The default
boot.type = "wild", Wald intervals and
varComprob fits are not affected.
The wild bootstrap (boot.type = "wild") and the
jackknife for method = "BCa" of lmer and
rlmer fits now use independent clusters when the model has
several grouping factors. Earlier versions used the grouping factor with
the most levels, which for nested designs such as
(1 | batch/cask) is the innermost one
(cask:batch): the wild bootstrap gave correlated casks
independent weights and understated the variability between batches, and
method = "BCa" failed and returned percentile intervals
with a warning. The clusters are now the coarsest grouping factor that
contains all the others (batch), the same rule as
simulate(type = "wild") in robustlmm. Models with crossed
grouping factors have no independent clusters and now stop with an error
unless clusterID is given. clusterID, the name
of a grouping factor or a vector with one value per observation, is no
longer ignored for lmer and rlmer fits with
boot.type = "wild" or method = "BCa". Models
with a single grouping factor give the same results as before (apart
from the BCa jackknife fix below).
The jackknife that estimates the acceleration of
method = "BCa" for lmer and rlmer
fits now leaves out one whole cluster at a time. Earlier versions split
the observations by the cluster sizes instead of by cluster, so each
jackknife sample left out a single observation. This changes all BCa
intervals for lmer and rlmer fits;
varComprob fits are not affected.