R help - Dec 2012 - kruskalmc, significant differences while median values are the same

Dear list!

I work with multiple Kruskal-Wallis test (kruskalmc, package pgirmess), which
evaluates differences in medians among groups (5 groups). A result of a test is
significant differences among some groups, while median values are the same for
4 groups (using tapply). Why?

p.s.: number of samples in groups vary from 50 to 4900.

Thanks to all, OV

.

Hello,

What about the median ranks? This test is based on ranks.

Regards,
Pascal


Le 12/12/15 18:04, Omphalodes Verna a ?crit :
> Dear list!
>
> I work with multiple Kruskal-Wallis test (kruskalmc, package pgirmess),
which evaluates differences in medians among groups (5 groups). A result of a
test is significant differences among some groups, while median values are the
same for 4 groups (using tapply). Why?
>
> p.s.: number of samples in groups vary from 50 to 4900.
>
> Thanks to all, OV
>
> .
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>

On Sat, Dec 15, 2012 at 10:04 PM, Omphalodes Verna <
omphalodes.verna@yahoo.com> wrote:
> Dear list!
>
> I work with multiple Kruskal-Wallis test (kruskalmc, package pgirmess),
> which evaluates differences in medians among groups (5 groups). A result of
> a test is significant differences among some groups, while median values
> are the same for 4 groups (using tapply). Why?
>
>
The Kruskal-Wallis test *doesn't* evaluate differences in medians, so there
is quite likely nothing wrong in a formal sense.

However, this does suggest that your groups may not be stochastically
ordered, which means the results of the Kruskal-Wallis test could be quite
misleading.  I'd suggest that you at least look at pairwise Wilcoxon tests
to make sure the direction agrees with what the Kruskal-Wallis test
implies. Box plots might also be a good idea.

Or, if you really want differences in medians, look at differences in
medians. A permutation test or a bootstrap confidence interval is probably
the best way to do this.


   -thomas

-- 
Thomas Lumley
Professor of Biostatistics
University of Auckland

	[[alternative HTML version deleted]]

Thank for help!

My ''problem'' is a little bit complicated. I have a dataset of
trees (five tree species) and I need to calculate if there are the significant
differences in the period of suppressed growth among tree species (length in
years, e.g. 1, 2, 3, 6, 10, 50, 80, etc.). Because data are not normally
distributed (Levene test), my idea was to use kruskalmc test.
?
So, my question is, how to evaluate the differences in the duration of
suppressed growth among groups? Is ''bootstrap confidence
interval'' for each tree species right solution??

thanks, OV




________________________________
From: Thomas Lumley <tlumley at uw.edu>
To: Omphalodes Verna <omphalodes.verna at yahoo.com> 
Cc: R Help <r-help at r-project.org> 
Sent: Sunday, December 16, 2012 11:59 PM
Subject: Re: [R] kruskalmc, significant differences while median values are the
same


On Sat, Dec 15, 2012 at 10:04 PM, Omphalodes Verna <omphalodes.verna at
yahoo.com> wrote:

Dear list!
>
>I work with multiple Kruskal-Wallis test (kruskalmc, package pgirmess),
which evaluates differences in medians among groups (5 groups). A result of a
test is significant differences among some groups, while median values are the
same for 4 groups (using tapply). Why?
>
>
The Kruskal-Wallis test *doesn't* evaluate differences in medians, so there
is quite likely nothing wrong in a formal sense.

However, this does suggest that your groups may not be stochastically ordered,
which means the results of the Kruskal-Wallis test could be quite misleading.
?I'd suggest that you at least look at pairwise Wilcoxon tests to make sure
the direction agrees with what the Kruskal-Wallis test implies. Box plots might
also be a good idea.

Or, if you really want differences in medians, look at differences in medians. A
permutation test or a bootstrap confidence interval is probably the best way to
do this.

? ?-thomas

-- 
Thomas Lumley
Professor of Biostatistics
University of Auckland?? ??