The list only includes bugs that were detected during work. Still active issues are on the top, solved issues are at the bottom.
When setting column names for a matrix the column names can themselves have names.
mat <- data.matrix(iris)
colnames(mat) <- c(A = "a", B = "b", C = "c", D = "d", E = "e")
colnames(mat)
# A B C D E
# "a" "b" "c" "d" "e"
The names are gone after any subset operation the matrix.
colnames(mat[1:10,])
# [1] "a" "b" "c" "d" "e"
Moreover, this is not the case for data.frames.
df <- iris
colnames(df) <- c(A = "a", B = "b", C = "c", D = "d", E = "e")
colnames(df)
# [1] "a" "b" "c" "d" "e"
R allows selecting elements either by index or by name. Having NA values in those selection are allowed and return an element with an unknown value - NA.
x <- c(a=1, b=2, c=3)
x[c(1,NA,3)] x[c("a", NA, "c")]
# a <NA> c # a <NA> c
# 1 NA 3 # 1 NA 3
But the above operations are inconsistent when working with matrices. Having NA selections when selecting by index works fine but the same operation throws an error when selecting by names.
mat <- matrix(c(1:6), ncol=3)
colnames(mat) <- c("a","b","c")
mat[,c(1,NA,3)] mat[,c("a", NA, "c")]
# a <NA> c # <error>
# [1,] 1 NA 5
# [2,] 2 NA 6
Function is.nan() doesn't work on data.frames while is.na() does.
is.na(iris) is.nan(iris)
# <works> # <error>
Autocompletion for list names does not work inside of square brackets.
iris$Sp<tab> iris[iris$Sp<tab>]
# works # does not work
Distribution functions like pnorm() have an inconsistent output format for an emtpy <0,0> dimension matrix.
x <- matrix(numeric(), nrow=0, ncol=0)
qbeta(x, 1, 1) # numeric(0)
qbinom(x, 1, 1) # numeric(0)
qbirthday(x, 1, 1) # <error>
qcauchy(x) # numeric(0)
qchisq(x, 1) # <0 x 0 matrix>
qexp(x, 1) # <0 x 0 matrix>
qf(x, 1, 1) # numeric(0)
qgamma(x, 1) # numeric(0)
qgeom(x, 1) # <0 x 0 matrix>
qhyper(x, 1, 1, 1) # numeric(0)
qlnorm(x) # numeric(0)
qlogis(x) # numeric(0)
qnbinom(x, 1, 1) # numeric(0)
qnorm(x) # numeric(0)
qpois(x, 1) # <0 x 0 matrix>
qsignrank(x, 1) # <0 x 0 matrix>
qsmirnov(x, c(1,1)) # numeric(0)
qt(x, 1) # <0 x 0 matrix>
qtukey(x, 1, 1) # numeric(0)
qunif(x) # numeric(0)
qweibull(x, 1) # numeric(0)
qwilcox(x, 1, 1) # numeric(0)
When the matrix is non empty all the functions (except qbirthday()) preserve the original matrix format.
Therefore, in my opinion, returning <0 x 0 matrix> is the correct behaviour.
When called on a data.frame with 0 columns rbind() drops all the rows, while cbind() keeps them.
This behaviour is also inconsistent with matrices, where both operations preserve the number of rows.
mat <- matrix(nrow=2, ncol=0) df <- as.data.frame(mat)
dim(mat) dim(df)
# [1] 2 0 # [1] 2 0
dim(cbind(mat, mat)) dim(cbind(df, df))
# [1] 2 0 # [1] 2 0
dim(rbind(mat, mat)) dim(rbind(df, df))
# [1] 4 0 # [1] 0 0
Note: this behaviour is documented.
When provided with all constant values some statistical tests silently produce NaN statistics and NA pvalues.
x <- c(1,1,1,1,1,1)
g <- c("a","a","a","b","b","b")
bartlett.test(x, g) # <NaN statistic, NA p-value, no warning, no error>
fligner.test(x, g) # <NaN statistic, NA p-value, no warning, no error>
kruskal.test(x, g) # <NaN statistic, NA p-value, no warning, no error>
oneway.test(x ~ g) # <NaN statistic, NA p-value, no warning, no error>
t.test(x ~ g) # <error>
var.test(x ~ g) # <NaN statistic, NA p-value, no warning, no error>
wilcox.test(x ~ g) # <warning + NA p-value>
Another case to check - when the values are constant within groups but differ between the groups
x <- c(1,1,1,2,2,2)
g <- c("a","a","a","b","b","b")
bartlett.test(x, g) # <NaN statistic, NA p-value, no warning, no error>
fligner.test(x, g) # <NaN statistic, NA p-value, no warning, no error>
kruskal.test(x, g) # <OK>
oneway.test(x ~ g) # <NaN statistic, NA p-value, no warning, no error>
t.test(x ~ g) # <error>
var.test(x ~ g) # <NaN statistic, NA p-value, no warning, no error>
wilcox.test(x ~ g) # <OK>
Note: the desired behaviour, in my opinion, is for all the cases with silent NA values or errors is to produce a warning instead.
When a paired wilcox.test() is called with missing observatiosn in y it shows a misleading error about observations missing in x.
wilcox.test(c(1,2), c(NA_integer_,NA_integer_), paired=TRUE)
# <error about not enough 'x' observations>
The acceptable input range for confidence level parameter conf.level is not consistent between tests.
x <- rnorm(10)
y <- rnorm(10)
g <- gl(2, 5)
binom.test(1, 1, 1, conf.level=1) # <error>
cor.test(x, y, conf.level=1) # <OK>
fisher.test(g, g, conf.level=1) # <error>
t.test(x, y, conf.level=0) # <OK>
mantelhaen.test(g, g, g, conf.level=1) # <OK>
var.test(x, y, conf.level=0) # <error>
wilcox.test(x, y, conf.int=0.95, conf.level=1) # <error>
Paired version of wilcoxon.test(paired = TRUE) has tolerance issues when detecting ties.
wilcox.test(c(4,3,2), c(3,2,1), paired=TRUE)
# <warning about ties>
wilcox.test(c(0.4,0.3,0.2), c(0.3,0.2,0.1), paired=TRUE)
# <no warning>
Information showing if normal approximation was used is missing from the output of wilcox.test().
wilcox.test(rnorm(10), exact=FALSE, correct=FALSE)
wilcox.test(rnorm(10), exact=TRUE, correct=FALSE)
# different results, but same data and identical method description
Different versions of wilcox.test() treats infinite values differently.
wilcox.test(c(1,2,3,4), c(0,9,8,Inf))
# Inf is removed
wilcox.test(c(1,2,3,4), c(0,9,8,Inf), paired=TRUE)
# Inf is treated as a highest rank
In some cases flinger.test() can produce highly significant result whan all groups have constant values.
fligner.test(c(1,1,1,2,2,2), c("a","a","a","b","b","b"))
# p-value < 2.2e-16
Function any() does not work on a logical data.frame but works on a numeric data.frame.
any(data.frame(A=0L, B=1L)) any(data.frame(A=TRUE, B=FALSE))
# TRUE # <error>
My opinion - any() not working on a data.frame might be an intended behaviour.
However then it also should not be working on a numeric data.frame.
Having a logical function - any() work on numeric inputs but not logical inputs is not expected.