Skip to content

Use cor.test for multiple parameters?

4 messages · Kathleen Regan, Eduard Szöcs, Wayne Richter +1 more

Original
Eduard Szöcs
#
On 11/15/2012 01:49 PM, Kathleen Regan wrote:
Hi Kathleen,

I suppose 'correlation.r' is a file with R-Code written by your 
colleague (probably containing the definition for the function 'cor_test()'.
If you don't have this file anymore, it wouldn't be hard to write some 
R-Code doing what you want (perhabs ~15Lines of Code).


Eduard
1 day later
Wayne Richter
# 
Take a look at rcorr() in the Hmisc package. You can divide the P value matrix by the value needed for the Bonferroni correction.

But, with 26 parameters, you will have 325 pairwise correlations. By dividing, for example, an alpha of 0.05 by 325, your minimum P value for significance would be 0.00015. It's going to take a mighty high correlation coefficient to be significant with 60 observations. Take a look at p.adjust.
From: 	Kathleen Regan <kath.regan at gmail.com>
To:	<r-sig-ecology at r-project.org>
Date: 	11/15/2012 8:16 AM
Subject: 	[R-sig-eco] Use cor.test for multiple parameters?

Please forgive me if this question has been answered somewhere else, but
for the life of me I couldn't find an answer (although i have found that
others have asked similar questions.)

If anyone replies, PLEASE keep in mind that I am a soil ecologist with
absolutely no background or experience in object-oriented programming,
which is perhaps why I have not been able to take advantage of existing
help on this subject. If there are answers, I simply do not understand them!

*What I need to do:*
I am trying to generate a table of correlations (26 parameters in columns,
60 data rows.) I need Spearman's rho, the significance level p, and I need
to apply the Bonferroni correction. I also need to export it into Excel in
a nice matrix-y table format.

*What I am able to do:*
I am able to do parts of what I want: with cor() I can generate a Spearman
correlation (but it's very ugly and hard to read or export nicely).
With cor.test I can generate a result when I use only 2 parameters (eg.
soilMois and BD), but I need to produce a table with all parameters
compared to each other.

Ex:     result<-cor.test(AprCorKR$SoilMois, AprCorKR$BD,method="spearman",
adjust.method="Bonferroni")

Warning message:

In cor.test.default(AprCorKR$SoilMois, AprCorKR$BD,  :

  Cannot compute exact p-values with ties


        Spearman's rank correlation rho

data:  AprCorKR$SoilMois and AprCorKR$BD
S = 49274.48, p-value = 0.003705
alternative hypothesis: true rho is not equal to 0
sample estimates:
       rho
-0.3691158

(By the way, I understand the warning message.)

*What I am unable to do:*
Use the function cor.test to generate the output of all my parameters with
respect to one another.

Examples below are of various ways of writing what I want to R:
adjust.method="Bonferroni")
Error: unexpected '[' in "result<-cor.test(AprCorKR$["
adjust.method="Bonferroni")
Error: unexpected '[' in "result<-cor.test(AprCorKR$["
adjust.method="Bonferroni")
Error: unexpected '[' in "result<-cor.test(c(["
adjust.method="Bonferroni")
Error in cor.test.default(vars = 2:26, vars = 2:26, data = AprCorKR, method
= "spearman",  :
  argument "x" is missing, with no default
adjust.method="Bonferroni")
Error in cor.test.default(vars = 2:26, data = AprCorKR, method =
"spearman",  :
  argument "x" is missing, with no default
adjust.method="Bonferroni")
Error in cor.test.default(data = AprCorKR, vars = 2:26, method =
"spearman",  :
  argument "x" is missing, with no default

*What I know:*
I have no missing values, no values of zero, and no typos that I am aware
of. I can display all my data correctly in R (and R seems to read it
correctly, based on what I am already able to do.)

A former colleague wrote a small line of code that does what I am trying to
do but I can't get it to work on my newer version of R. I have R 2.14.2 and
I don't know what version he was using.
His code for this: *result<-cor_test(data="**mydata*
",vars=2:26,adjust.method="Bonferroni")

(I include below his entire German instructions. But the code he wrote to
do this job is very short, and I would like to duplicate it if I can. It
provides the results I need in a form I can use.)

* *An additional note:
I have multtest loaded, and in my version cor.test is not written as
cor_test with an underscore but rather with a period. In the version of R
that I am using, there is no such thing as "correlation.r".

*
*


("Alb_Gesamtdat_10-03-10_bd_NA.txt") und das R-Skript ("correlation.r")
liegen und das Paket 'multtest' ist installiert
(www.r-project.org<file:///C:/Users/Regan/Documents/SCALEMIC/NewRStats/www.r-project.org>->
CRAN

-> Mirror -> Packages)
adjust.method="Bonferroni" )



Es werden bei diesem Funktionsaufruf automatisch 3 Dateien angelegt:

  - spearman_correlation_coefficients_complete_data.txt

  - spearman_correlation_raw_pvalues_complete_data.txt

  - spearman_correlation_Bonferroni_corrected_pvalues_complete_data.txt





Thanks to anyone who can help me with this!


Kind regards,

Kathy Regan
3 days later
Martin Weiser
# 
Hello Kathleen,
disclaimer: here I doubt Your approach, so if You are really sure what
You want, do not take this mail into account. Otherwise: 
I know nothing about your mental abilities, but so many pairwise
correlations are about order of magnitude more than I am able to handle
with mine. For me, this is a good reason to use some multivariate method
(e.g. ordination) and make things much more simple. Then test just what
seems to be meaningful (and correct for non-independence) if You really
need it. This way Your tests could be viewed as post-hoc ("325
tests"-way is rigorous, but I am not sure whether it is useful).  Or
forget testing, and just look at the matrix of correlation coefficients
themselves - these numbers tell the story, p-values just say how likely
are they lying in doing so.
Best,
Martin Weiser
 
Wayne Richter p??e v P? 16. 11. 2012 v 12:59 -0500: