New features
- Ordered factors can now be transformed to a finite collection
of order-restricted scores (see ?zheng_trafo); this can be used
to formulate score-independent tests for ordered categorical
data (see ?malformations, ?jobsatisfaction and ?vision).
- The test size, i.e., the actual significance level, of
approximative (Monte Carlo) and exact tests can now be computed
(see ?size).
- The centered linear statistic can now be extracted (see
?statistic).
- P-values can now be extracted from objects of class "PValue"
using pvalue().
Bug fixes
- Adjusted p-values based on the approximative (Monte Carlo)
marginal null distributions were incorrect under some
circumstances.
- The point and/or interval estimate of the ratio of scales
parameter (for tests against scale alternatives) were incorrect
when using the asymptotic null distribution.
- Exact distributions of symmetry tests for paired data, e.g.,
the exact sign and Wilcoxon signed-rank tests, used the
Streitberg-Röhmel algorithm for independent samples instead of
the more efficient version for dependent samples.
- The class definition of "PValue" erroneously included the
midpvalue and pvalueinterval slots; these slots are now part of
the "NullDistribution" subclass (as they should have been
originally).
User-visible changes
- Asymptotic and exact p-values smaller than machine precision,
i.e., .Machine$double.eps, or approximative (Monte Carlo)
p-values smaller than resampling precision, i.e., 1 /
nresample, are now reported as such in the printed output.
- Computations on the approximative (Monte Carlo) null
distribution, e.g., adjusted p-values for maximum-type tests,
are now more efficient.
- The Pearson chi-squared test is now computed more efficiently.
- Extracting the test statistic or the (global) p-value returned
a named result under some circumstances; the results are now
unnamed.
- The quantile function now returns a named vector if its p
argument is named.
- Objects of class "ApproxNullDistribution" now has a slot
nresample containing the number of Monte Carlo replicates.
- Classes "ExpectCovar" and "ExpectCovarInfluence" have been
moved to the party package.
- The ‘coin_implementation’ vignette has been renamed to
‘Implementation’.
- Documentation updates.
Other changes
- The linear statistic, its conditional expectation and
covariance, and permuted linear statistics are now computed by
package libcoin.
- confint() is now an S4 method.
- R >=3.4.0 and packages stats4 and matrixStats are now required.
Deprecated and defunct
- The B argument of approximate() and ApproxNullDistribution() is
now deprecated and has been replaced by nresample (for
consistency with the libcoin, party and partykit packages).
- Class "PValue" is now deprecated.