hlmLab 0.2.0
Correctness
hlm_context() now computes the standard error of the
contextual contrast from the full fixed-effect covariance matrix,
Var(bB - bW) = Var(bB) + Var(bW) - 2 Cov(bB, bW), instead
of assuming that the covariance is zero. In a random-intercept model
with an exactly group-mean centered Level-1 predictor that covariance is
zero by construction, so previously published random-intercept results
are unchanged; it is generally nonzero when the model contains a random
slope, when the predictor is centered in some other way, or when the
Mundlak specification uses the raw predictor.hlm_context() gains level and
check_centering arguments, returns conf_low
and conf_high columns, and stores the covariance of the two
coefficients in the attribute "cov_within_between". A
warning is issued when x_within does not appear to be
group-mean centered, because the difference between the two coefficients
is then not the contextual effect.hlm_icc() warns when the model has more than one
grouping factor, and when the first random-effect term contains a random
slope, in which case the reported ICC is conditional on the Level-1
predictor being zero.
Terminology
hlm_xint_geom() is deprecated. A random slope is
unexplained heterogeneity in a Level-1 association; it is not a
cross-level interaction, which requires an observed Level-2 moderator.
The function remains available as an alias that warns and forwards to
hlm_random_slope_plot(), and will be removed in a future
release.
New functions
hlm_random_slope_plot() displays cluster-specific
fitted lines from a random-slope model with the average line overlaid.
Clusters are selected evenly across the distribution of conditional
slopes by default, the plotting range is taken from the observed
predictor, and remaining fixed effects are held at their sample
means.hlm_cross_level_plot() displays the Level-1 association
at selected values of an observed Level-2 moderator, and warns when the
model contains no such interaction.hlm_icc_demo() simulates clustered data at several
target intraclass correlations with the total variance held constant, so
that students can see what low, moderate, and high ICC values look like.
Clusters are ordered by their mean within each panel
(sort_clusters = TRUE), so the sequence of cluster means is
nearly flat at a low ICC and steep at a high one.hlm_shrinkage_plot() compares raw cluster means with
multilevel empirical Bayes estimates and draws an arrow between them,
making partial pooling and its dependence on cluster size visible.
Other changes
hlm_icc() accepts cluster_size = "auto",
recovers cluster sizes from the fitted model, and reports an
unequal-cluster-size design effect based on
sum(n^2) / sum(n) alongside the usual mean-size
approximation. The returned object gains deff_unequal,
n_clusters, and mean_cluster_size.hlm_icc_plot() labels the components with their
percentages and reports the unequal-size design effect when
available.- A
testthat suite was added, including cases in which
the within- and between-cluster estimates are deliberately
correlated.
- Examples for
hlm_decompose() and
hlm_decompose_long() are self-contained and run without
external data.
hlmLab 0.1.0