NEWS | R Documentation |
Precision problem in the quantile function of exact null distributions computed by the shift algorithm.
The Gaugler-Kim-Liao class of tests for right-censored data is now
available (see ?logrank_test
and ?logrank_weight
).
The pvalue()
, midpvalue()
and pvalue_interval()
methods for objects of class "NullDistribution"
are now vectorized.
Precision problem in the density function of approximative (Monte Carlo) null distributions. This also impacts mid-p-values and p-value intervals.
Precision problem in the support of approximative (Monte Carlo) null distributions.
Precision problem in the symmetric median scores transformation.
The weighted logrank transformation returned nonsense scores when supplied non-right-censored input; it now results in an error.
The maximally selected statistics transformation for ordered factors sometimes returned the wrong labels.
The ties.method
argument of logrank_trafo()
no longer
allows using "logrank"
and "HL"
to specify mid-ranks and the
Hothorn-Lausen method respectively.
Documentation updates.
surv_test()
is now defunct and has been removed.
method = "discrete"
in pvalue()
is now defunct and has
been removed.
The dperm()
, pperm()
and qperm()
methods for
objects of class "AsymptNullDistribution"
are no longer needed and
have been removed.
Objects of class "AsIs"
were handled incorrectly in some
cases, leading to, e.g., misleading error messages or test statistics
having the wrong sign.
Supplying decreasing scores for ordered factors with two levels
using the test functions' scores
argument resulted in increasing
normalized scores, leading to test statistics having the wrong sign.
The transformation function for ordered factors did not return
normalized scores in the two-level case when argument scores
was
used.
Precision issues in confidence intervals for location and scale problems.
When the general symmetry and marginal homogeneity tests were
applied to an object of class "table"
, the levels of the
measurement conditions were not preserved (but reordered alphabetically).
Version 1.0-5 of the mvtnorm package is now required.
Documentation updates.
of_trafo()
did not handle ordered factors of length one.
Parallel operation (via approximate()
) changed the random
number generator state upon first invocation unless package
parallel's namespace had already been loaded.
Documentation updates.
Methods initialize()
and show()
were not imported from
package methods.
Documentation and vignette updates.
The Fisher-Yates correlation test (based on van der Waerden scores)
is now available (see ?fisyat_test
).
The quadrant test is now available (see ?quadrant_test
).
The Koziol-Nemec test and its corresponding transformation function
are now available (see ?koziol_test
and ?koziol_trafo
).
The Savage test and its corresponding transformation function are
now available (see ?savage_test
and ?savage_trafo
).
The Taha test is now available (see ?taha_test
).
The Klotz test and its corresponding transformation function are now
available (see ?klotz_test
and ?klotz_trafo
).
The Mood test and its corresponding transformation function are now
available (see ?mood_test
and ?mood_trafo
).
The Conover-Iman test is now available (see ?conover_test
).
The weighted logrank test and its corresponding transformation
function are now available, including weights for generalized Wilcoxon
tests (Gehan-Breslow, Prentice, Prentice-Marek and
Andersen-Borgan-Gill-Keiding), the Tarone-Ware class of tests, the
Fleming-Harrington class of tests and the Self class of tests (see
?logrank_test
, ?logrank_trafo
and ?logrank_weight
).
The sign test for paired data is now available (see
?sign_test
).
The Quade test is now available (see ?quade_test
).
The Pearson chi-squared test, the generalized
Cochran-Mantel-Haenszel test and the linear-by-linear association test now
allow the alternative hypothesis to be specified in the doubly ordered
case, i.e., when both the response variable and the explanatory variable
are ordered factors (see ?chisq_test
, ?cmh_test
and
?lbl_test
).
The Kruskal-Wallis test, the van der Waerden test, the Brown-Mood
median test, the logrank test and the Friedman test now allow the
alternative hypothesis to be specified when the explanatory variable is an
ordered factor (see ?kruskal_test
, ?normal_test
,
?median_test
, ?logrank_test
and ?friedman_test
).
The marginal homogeneity test now handles ordered responses and/or
ordered measurement conditions and allows the alternative hypothesis to be
specified in the doubly ordered case (see ?mh_test
).
The Brown-Mood median test and its corresponding transformation
function now offer the choice of three different versions of the mid-score
(see ?median_test
and ?median_trafo
).
The van der Waerden test, the Brown-Mood median test and the
Ansari-Bradley test now handle K-sample problems (see
?normal_test
, ?median_test
and ?ansari_test
).
The Fligner-Killeen test now handles two-sample problems (see
?fligner_test
).
Maximally selected statistics can now be computed for objects of
class "table"
(see ?maxstat_test
).
symmetry_test()
has gained a new argument paired
that
is used to indicate if paired data have been transformed in such a way
that the resulting unstandardized linear statistic is the sum of the
absolute values of the positive differences between the paired
observations (see ?symmetry_test
).
All transformation functions are now passing through missing values.
New function for rank transformations (see ?rank_trafo
).
New function for maximally selected statistics transformation of
ordered factors (see ?ofmaxstat_trafo
).
The Conover-Salsburg transformation function now allows the a
constant to be set (see ?consal_trafo
and ?neuropathy
).
The transformation function for ordered factors has gained a new
argument scores
that provides an alternative way of attaching
scores to ordinal variables (see ?of_trafo
).
The approximative (Monte Carlo) null distribution can now be
obtained by parallel operation using package parallel (see
?approximate
).
The exact null distribution of test statistics based on quadratic forms can now be obtained in the univariate two-sample case.
The exact null distribution can now be obtained for tests specified with case weights, implying that exact p-values can be computed for 2 x 2 tables or K x 2 tables with K ordered categories.
Evaluation of the null distribution can now be suppressed for all
test procedures by setting distribution = "none"
.
The mid-p-value can now be computed and is accompanied by a 99
% mid-p confidence interval when resampling has been used to obtain
the null distribution (see ?midpvalue
).
The p-value interval can now be computed (see
?pvalue_interval
).
Step-down adjusted p-values based on the joint distribution of the test statistics can now be obtained when the asymptotic null distribution is used.
Adjusted p-values based on the marginal distributions of the test statistics can now be obtained when the asymptotic null distribution is used.
Adjusted p-values based on the marginal distributions of the
test statistics now use a max-T procedure, instead of a min-P
(which is now deprecated), incorporating discrete distributional
characteristics when the approximative (Monte Carlo) null distribution is
used (see ?pvalue
).
Single-step or step-down adjusted p-values based on the
marginal distributions of the test statistics can now be obtained using
the Bonferroni or the Šidák method (see ?pvalue
).
Unadjusted p-values can now be obtained (see ?pvalue
).
Random numbers can now be generated from the permutation
distribution (see ?rperm
).
New data sets on gastric cancer, congenital sex organ
malformation and unaided distance vision have been added (see
?GTSG
, ?malformations
and ?vision
).
The standardized linear statistic of the Pearson chi-squared test was off by a factor sqrt(n / (n - 1)).
The Pearson chi-squared test ignored user-specified transformations.
Partial matching of the scores
argument did not work for
Pearson's chi-squared test and the marginal homogeneity test.
The Fisher-Pitman permutation test, the Wilcoxon-Mann-Whitney test, the Kruskal-Wallis test, the van der Waerden test, the Brown-Mood median test, the Ansari-Bradley test, the Fligner-Killeen test, the Wilcoxon signed-rank test, the Friedman test and the Page test did not return an error message when supplied with a censored response variable.
The Brown-Mood median test p-value computation used the wrong tail of the null distribution in the one-sided case.
The test statistic of the Wilcoxon signed-rank test sometimes had the wrong sign, due to being based on the sum of the absolute values of the negative, instead of the positive, differences between the paired measurements.
The Friedman test did not warn when weights were supplied.
Maximally selected statistics did not handle user-specified scores for ordinal variables correctly.
Two-sample problems were not identified as such if the grouping
variable had unused factor levels, causing problems if, e.g., trying to
use the subset
argument of formula
methods.
Stratified tests were not allowed if block
had unused factor
levels, causing problems, e.g., if trying to use the subset
argument of the formula
methods.
Transformations involving ordered factors or survival data sometimes failed when performed within blocks.
The maximally selected statistics transformation for factors sometimes dropped its column names.
The maximally selected statistics transformation for factors retained unused levels.
Approximative (Monte Carlo) p-values for stratified tests were wrong when the data were not ordered with respect to the blocks.
Exact p-values for stratified tests were wrong when each block contained precisely two observations.
Exact p-values for tests with an ordinal grouping variable were wrong.
The smallest single-step and step-down adjusted p-values sometimes differed from the global p-value.
Single-step adjusted p-values were too small when the support of the null distribution contained values equal to the observed test statistics.
Monotonicity of step-down adjusted p-values was not enforced when using the approximative (Monte Carlo) null distribution.
The density function of asymptotic permutation distributions for
maximum-type tests returned nonsense; it now reports NA
.
The density function of approximative (Monte Carlo) permutation distributions returned nonzero probabilities for values not included in the support.
The density function of exact permutation distributions returned a zero-length vector for values not included in the support; the probability of such values are now reported as 0.
The distribution and quantile functions of asymptotic permutation distributions for maximum-type tests were not handling vector arguments correctly.
The quantile function of exact permutation distributions sometimes
reported NA
when the 100 % quantile was requested.
The support of exact permutation distributions for stratified tests was not guaranteed to be distinct and ordered, which in turn caused failure when trying to obtain values from the density and quantile functions.
The printed output of the Pearson chi-squared test and the generalized Cochran-Mantel-Haenszel test claimed that a Linear-by-Linear Association Test had been carried out even in the singly ordered case, i.e., when either the response variable or the explanatory variable is an ordered factor; these are now reported as Generalized Pearson Chi-squared Test and Generalized Cochran-Mantel-Haenszel Test respectively.
The printed output of the Fisher-Pitman permutation test reported the parameter tested in the two-sample case when the explanatory variable was an ordered factor.
The printed output of the one-sided Ansari-Bradley test reported the opposite direction of the specified alternative (but performed the test in the specified direction and thus reported the correct p-value).
The printed output of the logrank test did not report whether a two- or K-sample test had been carried out.
The printed output of the logrank test did not report the tested parameter in the two-sample case.
The printed output of maximum-type tests did not include the alternative hypothesis.
The printed output of the general symmetry test claimed that a General Independence Test had been carried out; it is now reported as General Symmetry Test.
Objects of class "SymmetryProblem"
were not checked for
validity.
The Pearson chi-squared test, the generalized Cochran-Mantel-Haenszel test, the linear-by-linear association test and the marginal homogeneity test now use a scalar test statistic in the doubly ordered case.
The Fisher-Pitman permutation test, the Kruskal-Wallis test, the logrank test and the Friedman test now use a scalar test statistic in case the explanatory variable is an ordered factor.
The Fisher-Pitman permutation test no longer allows the test statistic to be specified; a quadratic form is now used in the K-sample case.
The Pearson chi-squared test, the generalized
Cochran-Mantel-Haenszel test, the Kruskal-Wallis test, the Friedman test,
the Fligner-Killeen test and the marginal homogeneity test no longer
features the distribution
argument (but still allow the type of
null distribution to be specified).
The Ansari-Bradley test no longer features the alternative
argument (but still allows the alternative hypothesis to be specified).
When the general symmetry test or the marginal homogeneity test is applied to a table, the variable indicating the measurement conditions has been renamed to conditions (from groups).
The teststat
argument no longer allows the use of
maximum-type statistics and quadratic forms to be specified using
"maxtype"
and "quadtype"
respectively.
The median and logrank transformations are now increasing functions.
The default number of Monte Carlo replicates used to approximate the exact conditional distribution has been increased to 10000.
The exact distribution of stratified test statistics is now obtained more efficiently, and especially so for large numbers of strata.
exact()
has algorithm = "auto"
as its new default,
preserving the behavior of algorithm = "shift"
in previous versions
of the package.
The shift algorithm (exact(algorithm = "shift")
) no longer
silently switches to the split-up algorithm when non-integer scores are
detected; it now fails with an error message.
Single-step adjusted p-values based on the joint distribution of the test statistics are now computed slightly faster due to using only the unique realizations of the test statistics.
The support of an asymptotic null distribution is now reported as
NA
.
Small changes in the printed output of the Pearson chi-squared test, the Fisher-Pitman permutation test, the Wilcoxon-Mann-Whitney test, the van der Waerden test, the Brown-Mood median test, the marginal homogeneity test, the Wilcoxon signed-rank test, and generalized maximally selected statistics.
The printed output of the Fisher-Pitman permutation test now reports that a Linear-by-Linear Association Test is carried out when the explanatory variable is an ordered factor.
In the printed output of the Ansari-Bradley test and the Spearman test the tested parameter has been renamed to ratio of scales and rho respectively.
wilcoxsign_test.IndependenceProblem()
has been replaced by
wilcoxsign_test.SymmetryProblem()
.
The "IndependenceTestStatistic"
class is now virtual.
Objects of class "IndependenceProblem"
drop unused factor
levels.
Objects of classes "ScalarIndependenceTestStatistic"
or
"QuadTypeIndependenceTestStatistic"
now have a paired
slot.
Objects of class "ScalarIndependenceTest"
now has a
parameter
slot.
Objects of classes "ApproxNullDistribution"
or
"AsymptNullDistribution"
now have a slot seed
containing the
state of the random number generator.
The package manual has been completely revised.
The memory consumption during computations has been reduced.
Henric Winell, a long-term collaborator who donated almost all changes in coin 1.1-0, is now listed as author. Welcome on board, Henric!
surv_test()
is now deprecated and has been replaced by
logrank_test()
.
method = "discrete"
in pvalue()
is now deprecated.
The Wilcoxon signed-rank test no longer features the defunct
ties.method
argument.
as.integer(round(x))
is safer than as.integer(x)
due to
truncation in the latter.
Fix ‘DESCRIPTION’ and ‘NAMESPACE’ issues.
Documentation updates.
Test update (for Sparc Solaris).
Move ‘inst/doc/vignettes/*’ to ‘vignettes/*’.
New CWD
data.
Report a warning for rank tests with weights.
No require()
in .onLoad()
.
exact()
can not work for symmetry_test()
; spotted by LE
PAPE Gilles lepape.gilles@neuf.fr.
Add dependencies required for ‘tests/*’.
Change ties.method = c("HollanderWolfe", "Pratt")
to
zero.method = c("Pratt", "Wilcoxon")
in wilcoxsign_test()
following a suggestion by Fritz Scholz. A warning is fired to avoid
misleading and nonreproducible results.
Use quantile(..., type = 1)
for computing approximate quantiles
(suggested by Fritz Scholz).
Documentation Biobase::expressionSet
.
Further checks and better error messages for nonsense data in (paired) two-sample tests.
Better error message for
wilcoxsign_test(c(1, 1, 1) ~ c(1, 1, 1))
.
Bugfix in discrete MTP adjustment.
Improve upon qperm()
for van de Wiel algorithm again.
Again precision problems in van de Wiel algorithm: qperm()
suffered from too large tolerances of the inverted probability function.
Precision problems in van de Wiel algorithm fixed. In rare cases,
P(T < t) was returned as p-value instead of P(T ≤ t).
Two statistics are now considered equal (in all computations) when the
difference is smaller than sqrt(.Machine$double.eps)
.
dperm()
, pperm()
and qperm()
are vectorized also
for exact and approximative null distributions.
Exact distribution for independent two-sample problems with only two observations was wrong, spotted by Fritz Scholz fscholz@u.washington.edu.
Documentation updates.
Evaluate all formula
e in xxx_test()
's
parent.frame
.
Add alternative ties handling to wilcoxsign_test()
feature
request by Fritz Scholz fscholz@u.washington.edu.
Vignette update.
Fix location confidence interval problem spotted by Fritz Scholz fscholz@u.washington.edu.
Add average scores for logrank test.
Fix Rd problems.
JSS paper http://www.jstatsoft.org/v28/i08/ documents version 1.0-0.
Set default for <IndependenceTest>@method
to
General Independence Test.
covariance()
always returns a covariance matrix.
Add show()
method for "IndependenceTest"
objects.
Functions supplied via distribution
does not need to have a
class.
Export all classes.
Compute linear statistics, expectations and (co)variances when constructing <IndependenceLinearStatistic> instead of <IndependenceTestStatistic> and define methods for this class.
maxstat_test()
with integer weights gave NA
or wrong
cutpoints.
any()
was misused in one place, spotted by Kasper Daniel Hansen
khansen@stat.Berkeley.EDU.
Code cosmetics by Johannes Huesing johannes@huesing.name.
Remove non_function entries from Rd files.
Disable Biobase
example in vignette.
Fix precision problems with exact p-values (spotted by Michael Fay mfay@niaid.nih.gov.
Add new vignette on technical details.
Update to new mvtnorm 0.8-0.
Check for overflow errors in ‘StreitbergRoehmel.c’ (thanks for Michael Fay mfay@niaid.nih.gov for spotting this).
Add new argument ordered_trafo
to trafo()
and deal with
ordered factors in a more transparent way (via a new function
of_trafo()
).
New vignette ‘MAXtest’.
Print name of x
variable and levels when x
is a factor.
New class "IndependenceLinearStatistic"
.
LazyLoad: yes
.
Rename some C source files & update doxygen documentation.
Simplify some methods and the class structure.
Clean up code and simplify wrapper functions.
Documentation updates.
Deal with factors containing only one level.
Do not compute design matrices in ModelEnvFormula
.
Expand weights if distribution = approximate()
.
Assign names to user-specified transformation in case they are missing.
The distribution
argument may now be a function with one
argument allowing for user-specified distributions.
surv_test()
ignored the alternative
argument.
maxstat_trafo
did not always choose the correct maximal
cutpoint.
Formula evaluation without data
argument was partially broken
(thanks to Achim Zeileis for spotting this).
Improve formula interface for wilcoxsign_test()
.
Implement exact distribution for symmetry problems (especially
wilcoxsign_test()
).
Add more checks on StatXact examples.
Fix problems reported by valgrind.
‘coin-Ex.R’ generated by R 2.4.0.
Enhances: Biobase
.
Add $(FLIBS)
to ‘Makevars’.
Include doxygen documentation for C functions.
Documentation updates.
maxstat_test(y ~ x)
is now able to deal with unordered x
variables.
maxstat_test()
returns estimates of both the selected variable
and the cutpoint in multivariate situations (as a list).
More checks for maxstat_test()
.
Add codetools checks to tests.
Fix typo: Homegeneity.
Add ASA copyright to ‘LegoCondInf’ vignette.
Clarify that the Stuart-Maxwell test is computed by mh_test()
and add a further example (thanks to Henric Nilsson
henric.nilsson@phadia.com for pointing this out!).
Improve maxstat_trafo()
: some potential cutpoints could have
been overlooked in case of ties equal to the maxprob and minprob sample
quantiles.
Reproduce two examples from Hothorn and Lausen (2003) in
?maxstat_test
(and added hohnloser
data set).
Add more regression tests on maxstat_test()
.
Updated ‘LegoCondInf’ vignette.
Printed names of test statistics are now chi-squared, maxT, or Z instead of T.
User-supplied transformations can return a vector which is coerced to
matrix(..., ncol = 1)
internally.
ytrafo
and xtrafo
may take functions like rank()
directly. However, the use of trafo()
is recommended.
Argument teststat
may take values "max"
, "quad"
or "scalar"
.
Internal reimplementation of score handling for ordered factors.
Handling of multiple ordered factors / ordered factors in multivariate situations implemented.
f_trafo()
returns a design matrix with NA
rows in case
missing values were present.
Make R CMD check happy and move ‘src/README’ to ‘inst/README’.
isequal()
must not check equality of attributes.
New data set alpha
.
Data set alzheimer
is now a "data.frame"
, not a
"table"
.
New vignette with more applications.
Documentation updates.
Remove unused setAs
definitions.
expectation()
, variance()
and covariance()
return
named vectors or matrices. Names for objects returned by statistic()
have been partially improved.
All observations with NA
s are removed now.
Some internal improvements (removed unused code, avoid duplicated code chunks etc.).
More checks on NA
s, blocks and multiple ordinal variables (one
univariate problems are currently allowed to have ordinal variables).
Export independence_test.table()
.
Add conf.level
attribute for MCp
objects when
quasi-randomized Monte Carlo procedures (from mvtnorm) have been used.
Add new var_trafo
argument to trafo()
.
Documentation updates.
statistic()
now returns correct linear and standardized
statistics when scores are in play.
alzheimer
data added.
photocar
data added.
Documentation updates.
support()
and dperm()
methods have been improved for
asymptotic and approximative null distributions.
Internal C function R_MonteCarloIndependenceTest
returns a
pq x B matrix instead of a list with B elements (the
linear statistic for each random permutation of the data).
Various simplifications for the computations of adjusted p-values.
I()
in formula
e could still cause trouble with class
AsIs
, now fixed.
Logical variables are now allowed in formula
e (and are treated
like factor
s).
Fix print
method for marginal homogeneity tests
(...stratified by block...).
R_kronecker
in C now available.
C versions of nrow()
and ncol()
return 1 or length for
vectors.
Some internal optimisation.
C functions nrow(x)
and ncol(x)
return LENGTH(x)
or 1
when x
has no dim
attribute.
New R-interface function R_kronecker
to C_kronecker
(which returns a _vector_!).
Speedup of variance computations in internal functions.
Fix more problems reported by new codetools. Try to work around
the terms(y ~ ., data = data.frame(1:10))
problem in R 2.2.0.
Fix some problems reported by codetools.
alternative = "less"
and alternative = "greater"
are now
defined for "maxtype"
statistics as well.
One- and two-sided single-step and step-down max-T p-value
adjustments are now available from the appropriate pvalue()
method.
(NOTE: those procedures have not been tested carefully, yet.)
In addition, the Bonferroni-adjustment by Westfall & Wolfinger (1997) is
available now. Note that the interface to pvalue()
changed slightly,
adjustment = TRUE
was replaced by method = "single-step"
.
More examples added to the vignette.
distribution = "approximate"
for "maxtype"
statistics
was wrong in case both xtrafo
and ytrafo
were
multivariate (the conditional expectation was computed incorrectly).
I()
in formula
e returns objects of class AsIs
which caused troubles in trafo()
.
Functionality for formula
parsing and evaluation is now
imported from package modeltools.
show()
returns objects (of class "htest"
, for example)
invisibly, really.
maxstat_trafo()
is much faster now and returns a matrix with
both row- and column names set appropriately.
The distribution
argument now takes the return values of
functions exact()
, approximate()
or asymptotic()
as
well. Those functions can be used to specify parameters, such as the number
of Monte Carlo replications via
..., distribution = approximate(B = 9999), ...
.
show()
returns objects (of class "htest"
, for example)
invisibly.
expectation()
returns a vector, not a matrix.
New generic variance()
for extracting the variance(s) of linear
statistics.
Only variances (instead of the whole covariance matrix) is computed when the distribution of maximum type test statistics is to be approximated.
data
may be an object of class "exprSet"
(->
Biobase), the vignette has an example.
logrank_trafo()
(and surv_test()
) now have a
ties.method
argument, see ?surv_test
for more information.
asymptotical -> asymptotic in
print
methods.
mercuryfish
example added.
<x,y>trafo
now can return matrices with number of columns
different from the lhs and rhs of formula
.
mergesort
is already defined in
‘/usr/include/stdlib.h’:270 on some platforms.
delay
is deprecated in R 2.1.0.