Kruskal-Wallis test with details on pairwise comparisons

Question

The standard stats::kruskal.test module allows to calculate the kruskal-wallis test on a dataset:

>>> data(diamonds)
>>> kruskal.test(price~carat, data=diamonds)

Kruskal-Wallis rank sum test

data:  price by carat by color 
Kruskal-Wallis chi-squared = 50570.15, df = 272, p-value < 2.2e-16

This is correct, it is giving me a probability that all the groups in the data have the same mean.

However, I would like to have the details for each pair comparison, like if diamonds of colors D and E have the same mean price, as some other softwares do (SPSS) when you ask for a Kruskal test.

I have found kruskalmc from the package pgirmess which allows me to do what I want to do:

> kruskalmc(diamonds$price, diamonds$color)
Multiple comparison test after Kruskal-Wallis 
p.value: 0.05 
Comparisons
      obs.dif critical.dif difference
D-E  571.7459     747.4962      FALSE
D-F 2237.4309     751.5684       TRUE
D-G 2643.1778     726.9854       TRUE
D-H 4539.4392     774.4809       TRUE
D-I 6002.6286     862.0150       TRUE
D-J 8077.2871    1061.7451       TRUE
E-F 2809.1767     680.4144       TRUE
E-G 3214.9237     653.1587       TRUE
E-H 5111.1851     705.6410       TRUE
E-I 6574.3744     800.7362       TRUE
E-J 8649.0330    1012.6260       TRUE
F-G  405.7470     657.8152      FALSE
F-H 2302.0083     709.9533       TRUE
F-I 3765.1977     804.5390       TRUE
F-J 5839.8562    1015.6357       TRUE
G-H 1896.2614     683.8760       TRUE
G-I 3359.4507     781.6237       TRUE
G-J 5434.1093     997.5813       TRUE
H-I 1463.1894     825.9834       TRUE
H-J 3537.8479    1032.7058       TRUE
I-J 2074.6585    1099.8776       TRUE

However, this package only allows for one categoric variable (e.g. I can't study the prices clustered by color and by carat, as I can do with kruskal.test), and I don't know anything about the pgirmess package, whether it is maintained or not, or if it is tested.

Can you recommend me a package to execute the Kruskal-Wallis test which returns details for every comparison? How would you handle the problem?

Answer 1

One other approach besides kruskal::agricolae mentioned by Marek, is the Nemenyi-Damico-Wolfe-Dunn test implemented in the help page for oneway_test in the coin package that uses multcomp. Using hadley's setup and reducing the B= value for the approximate() function so it finishes in finite time:

#updated translation of help page implementation of NDWD
NDWD <- 
    independence_test(dv ~ iv, data = sum_codings1, distribution = approximate(B = 10000), 
                          ytrafo = function(data) trafo(data, numeric_trafo = rank_trafo), 
                          xtrafo = mcp_trafo(iv = "Tukey"))


    ### global p-value
    print(pvalue(NDWD))

    ### sites (I = II) != (III = IV) at alpha = 0.01 (page 244)
    print(pvalue(NDWD, method = "single-step"))

More stable results on that larger dataset may require increasing the B value and increasing the user's patience.

Jan: 2012: There was recently a posting on R-help claiming unexpected results from this method so I forwarded that email to the maintainer. Mark Difford said he had confirmed the problems and offered an alternate tests with the nparcomp package: https://stat.ethz.ch/pipermail/r-help/2012-January/300100.html

There were also in the same week a couple of other suggestions on rhelp for post-hoc contrasts to KW tests: kruskalmc suggested by Mario Garrido Escudero and rms::polr followed by rms::contrasts suggested by Frank Harrell https://stat.ethz.ch/pipermail/r-help/2012-January/300329.html

Nov 2015: Agree with toto_tico that help page code of coin package has been changed in the intervening years. The ?independence_test help page now offers a multivariate-KW test and the ?oneway_test help page has replace its earlier implementation with the code above usng the independence_test function.

Answer 2

You can use PMCMR package. There is more information about it.

Spelling_Grades <- c(90,87,89,90,75,88,97,99,78,85,72,76,77,79,70)
Methods <- c("A","A","A","A","B","B","B","B","B","B","C","C","C","C","C")
kruskalmc(Spelling_Grades~Methods)

#This method doesn't accept characters that's why I've changed the methods to integer
Methods <- c(1,1,1,1,2,2,2,2,2,2,3,3,3,3,3)
posthoc.kruskal.nemenyi.test(Spelling_Grades~Methods) 

The two methods above give same results.

Answer 3

I would have thought you'd be able to do the following:

data(diamonds, package = "ggplot2")

library(coin)
library(multcomp)

kt <- kruskal_test(price ~ clarity, data = diamonds)
glht(kt, mcp(clarity = "Tukey"))

But it seems like multcomp doesn't support coin objects (yet?).

Answer 4

Unfortunately I don't know of a function like this. If there isn't one already, it would be an interesting task to construct a function that returns a matrix with all pairwise treatment comparisons. A contrast is considered significant if the following inequality is satisfied

_{(source: statsdirect.com)}

where T is the Kruskal-Wallis test statistic for k samples, S^2 is the denominator of the T statistic, N is the total number (all ni) and Ri is the sum of the ranks (from all samples pooled) for the ith sample, and t is a quantile from the Student t distribution on N-k degrees of freedom.

I know I didn't help much :)
I am also waiting for a better answer