[R] Theil test help

David Winsemius dwinsemius at comcast.net
Fri Jul 17 00:49:59 CEST 2009


On Jul 16, 2009, at 5:53 PM, KARSTEN G HOLMQUIST wrote:

> Hello,
>
> I have a series of questions that I hope will be simple to answer.   
> Basically I would like a code to do the following so that I can  
> compute the distribution free test for the slope of a postulated  
> regression line (Theil test).  As I am testing the null hypothesis  
> that slope = 0 against the general alternative the slope does not  
> equal 0, it should be pretty straight-forward.
>
> I have a data frame as follows;
>
>
> index	Score	Mean_Temp	SD_Mean_Temp
> i	10	3.7		0.65
> i+1	12	7.5		2.11
> i+2	6	10.3		0.85
> ...	...	...		...
> n	8	12.6		4.2
>
> The first question is how can I compute the n(n-1)/2 differences  
> between Score i and Score j where i is from 1 to (n-1) and j is from  
> (i+1) to n?  I would then like R to spit out the differences in a  
> file something like the following;
>
> i	j	Difference
> 1	2	2
> 1	3	-4
> ...	..	...
> (n-1)	n	D(n-1,n)

Without an example, I am going to suggest something:

dfg<-data.frame(index=1:10, Score=sample(1:20,  
10),Mean_Temp=10*abs(rnorm(10)), SD_Mean_Temp=abs(rnorm(10)) )
>
The trick is to get the differences and here is one method:

mapply(FUN="-", dfg$Mean_Temp[combn(dfg$index,2)[1,] ], dfg 
$Mean_Temp[combn(dfg$index,2)[2,] ])

Should be pretty obvious how to line up the indices since they are  
just the arguments to "[".



> The second question is for each indexed sample from i to n, I would  
> like to use something like rnorm(n, mean = 0, sd = 1) so that I  
> could generate 1000 random draws from the distribution specified so  
> that the arguments for rnorm () are as follows; mean = the value of  
> Mean_Temp & sd = SD_Mean_Temp for indexed values i to n.  I would,  
> of course, then like R to spit out the massive table with 1000  
> columns of randomly generated temperature values for each index i to  
> n.

You can get 1000 such values with rnorm(1000, Mean_Temp, SD_Mean_Temp)

Perhaps something along these lines:

Mtx <- matrix( , ncol=1000, nrow=nrow(dfg))
Mtx[1:nrow(dfg), ] <- apply(dfg, 1, function(x) rnorm(1000, x[3],  
x[4]) )


> Finally I would like to compute the slope estimator associated with  
> the Theil statistic.  All I need to do is compute the n(n-1)/n
                            ^^^^^^^^^^==n,    is that what you meant?  
Or is it n(n-1)/2?

> individual slope values; Sij = (Yj -Yi)/(Xj-Xi) where the Ys are the  
> Scores for samples i to n and the Xs are the set of randomly  
> generated temperature values, as before, 1 less than or equal to i  
> less than j less than or equal to n.  The median is then the  
> estimator of Beta (the slope).  I would like to compute the slopes  
> for all 1000 sets of randomly generated temperature values.

There is a package, mblm, that computes the Theil estimator. I don't  
see why we should reinvent that wheel.

(Found via tsearching on theil statistic)
http://search.r-project.org/cgi-bin/namazu.cgi?query=theil+statistic&max=100&result=normal&sort=score&idxname=functions&idxname=Rhelp08&idxname=views
...and following what seemed like promising links


library(mblm)

-- 

David Winsemius, MD
Heritage Laboratories
West Hartford, CT




More information about the R-help mailing list