wilcoxon sign ranked test

[seanpps2]

hi i am trying to do a program which does a wilcoxon signed ranked test . However i discover that the result from labview differs from that for matlab . Is that some possible reason ?

Data Used 

A : 5 6 3 6 7 2 5 6 8

B : 8 6 0 7 4 5 9 3 7  

labview p value 0.0468

matlab p value 0.9036

Thank you.

[GerdW]

Hi seanpps2,

don't you think we could evaluate your example much better/easier if you would attach both the VI and the m-code you used to get those rank numbers?

[seanpps2]

Hi GerdW,

I have attached the file as requested . As i am unable to attach the . file the code is as follows ,

x = [ 5 6 3 6 7 2 5 6 8];
y = [8 6 0 7 4 5 9 3 7 ];
[p, h] = signrank(x,y,'alpha',0.05);
disp(p);
disp(h);

Thank you .

I have made a mistake regarding the p values it should be 0.9219 -> matlab , 0.7421 -> labview . 

[Alisha_P]

Hi seanpps2,

It does seem strange that these two pieces of software get different answers.

Do you know which one is the answer you expect?

I have submitted this issue to R&D for investigation.

[seanpps2]

Hi Alisha, 

Unfortunately i do not have a standard answer for the stats. test. Do you have any idea of a posible timeline with regards to the problem as we are looking to integrate this into our product. Thank you for your assistance.

[GerdW]

Hi sean,

maybe you could provide a dataset with known results? Then a comparison of LV vs Matlab would be easier...

[Alisha_P]

Hi sean,

At this point, we don't have much knowledge about what is going on, whether there is anything wrong with this calculation or not so we don't have any timeline at this point. 

A dataset with known results would indeed be helpful.

[seanpps2]

Hi , 

I have tested with several stats. software such as SPSS and miniTab . It seems that for the above mentioned dataset different result are obtained. That why it is rather hard to provide a sample dataset with a known value . 

For instance for the below dataset 

X : 12 13 15 19 20 21 27

Y : 18 23 24 30 32 35 40

BOTH labview and matlab obtained a result of 0.0156. However , this dataset is obtained from a college textbook(reliable source i believed) , Statistic for Engineer and Scientist (William Navidi) and it shows an answer of 0.0174.

Thank you .

[Darin.K]

There are some serious flaws in the LV implementation.  Your last dataset is immune because the values for the differences are unique.  The serious issues with the LV version include:

1.  The rank of multiple ties is averaged, all zeros should be ignored.  Instead the rank of two ties are each 0.5, the rank of three ties is 2.  This error propogates through the remaining rank values.  The lowest non-tied value should have rank 1.

3.  The ranks are converted to I32 after the Rank Transformation VI, this rounds off the 0.5 values from ties and again gives erroneous results.

4.  For some reason the rank is incremented after it is converted to I32, this probably worked in some special case, but not in general.

5.  In the lookup table (for small samples like this), the index array should probably be Interpolate 1D array instead.  In cases with ties, you get a better match.

With these fixed I get reasonable, but not perfect agreement with Matlab (I get p=0.8945 for your first dataset).  Mathematica gives me slightly different results still, actually closer to the LV value (p=0.8854). 

[seanpps2]

Hi Darin , 

Thank you for your explaination. Do you mind if you share your implementation ?

Regards