Raw.Significant is now derived from a comparison
against a randomly generated variable’s weight, rather than from the
confidence interval around the raw weight itself (#26). Raw relative
weights are non-negative, so an interval around a weight almost always
excludes zero and previously flagged even unrelated predictors as
significant. Comparing each weight to that of a random variable is the
approach suggested by Tonidandel, LeBreton and Johnson (2009, https://doi.org/10.1037/a0017735) for judging whether a
weight exceeds what chance alone would produce (see also the discussion
in vignette("evaluating-rwa-method-reference")). This
package applies a directional cutoff on that comparison: a predictor is
significant only when the lower bound of the difference interval is
above zero, so a predictor that performed worse than the random variable
is not reported as significant either. Predictors previously
reported as significant may now correctly be reported as not
significant. rwa() also returns
Random.Diff.CI.Lower and Random.Diff.CI.Upper
for the comparison the flag is based on, and the descriptive
Raw.RelWeight.CI.* columns are unchanged.bootstrap = TRUE now runs an additional bootstrap when
comprehensive = FALSE, which roughly doubles bootstrap
time. comprehensive = TRUE already computed the comparison
and is unaffected.rwa_logit() and rwa_multiregress()
to support logistic regression and multiple regression.use parameter to rwa() function to
control how missing data is handled when computing correlations. Options
include “pairwise.complete.obs” (default, pairwise deletion),
“complete.obs” (listwise deletion), and other standard options from
cor(). (#12)weight parameter to rwa() function
to perform observation-weighted Relative Weights Analysis (#12), using a
weighted complete-case correlation matrix. Bootstrap inference uses iid
individual-row resampling, carrying each row’s weight; clusters, strata,
and replicate-weight survey designs are not supported.rwa() and
rwa_multiregress() now include n_weighted (sum
of retained original weights) and n_effective (Kish’s
unequal-weighting effective sample size). Existing unweighted return
fields are unchanged.use and weight arguments, the weighted
sample-size diagnostics, and the multiple-regression-only scope.
Corrected the introductory vignette’s incorrect statement that missing
data is handled by listwise deletion; the default has been pairwise
deletion.use
and weight parametersall.obs;
corrected na.or.complete documentation.plot_rwa() now reports the sum of weights and the
effective sample size in the caption for weighted analyses, so weighted
charts are distinguishable from unweighted ones.n_weighted and
n_effective immediately after n, making them
easier to find. Field names and unweighted output are unchanged.?rwa, including guidance on when to use survey
weights.rand.rwa() parameters including conf_level,
n_bootstrap, non-numeric variables, zero-variance
variables, and singular correlation matrices with informative error
messagesbootstrap = TRUE parameter to rwa() for
statistical significance testing of relative weightssort = TRUE
parameter to automatically sort results by importance (descending
order). Set sort = FALSE to preserve original predictor
orderAuthors@R field for CRAN submissionboot,
purrr, and utils packages for bootstrap
functionalityRe-submission to CRAN
\donttest{} in examples where unnecessaryRe-submission to CRAN
First submission to CRAN (required to re-submit)
rwa()plot_rwa()remove_all_na_cols()%>% operator is exported