is it correct if I use Mann-Whitney U?

[Omar Farook]

Dear Experts,
  I have a 72 reading , the first 36 reading are from tooth one and  the rest from tooth two , the readings are from 36 patient , i.e. two readings per patient.
  The readings format are 0 , 1 , 2 , and 3 , and each value representing rang, for example , 0 representing from 0 -9.
  I want to test if there is a significant difference between the two readings of each patient.
  So by using SPSS , is it correct if I use Mann-Whitney U?
  Many thanks in advance.
  Omar.


---------------------------------
Need a quick answer? Get one in minutes from people who know. Ask your question on Yahoo! Answers.

[Simon, Steve, PhD]

Omar writes:

> I have a 72 reading , the first 36 reading are from tooth one
> and the rest from tooth two , the readings are from 36 patient ,
> i.e. two readings per patient. The readings format are 0 , 1 , 2 ,
> and 3 , and each value representing rang, for example , 0
> representing from 0 -9. I want to test if there is a significant
> difference between the two readings of each patient. So by using
> SPSS , is it correct if I use Mann-Whitney U?

I'm not quite sure what the word "rang" means, so perhaps you can
clarify this. But I am quite certain that you do not want the
Mann-Whitney U. This statistic compares two independent groups, and your
groups are clearly related. Two teeth that come from the same mouth
share a common environment and are likely to be strongly correlated. In
fact, you're glad they are strongly correlated, because it improves the
efficiency of your analysis.

If you look at the SPSS menu for nonparametric tests, there is a choice
for "two related samples." This brings up a dialog box which offers the
Wilcoxon, Sign, and McNemar tests. I suspect you want the McNemar test.
Wilcoxon isn't going to work so well because you have a lot of ties in
your data set, and the sign test, though conceptually very simple and
with very few assumptions, just isn't quite as efficient.

If the code 0 represents a range (rang?) of values then you might be
able to gain even more efficiency by using the original values rather
than summarizing them as inside certain ranges. Without more context,
though, I can only speculate.

Finally, if your data is a single column rather than two adjacent
columns, you may need to use the restructure data wizard.

I hope this helps.

Steve Simon, [hidden email], Standard Disclaimer
My niece (Bachelor's degree in Chemical Engineering
from Carnegie Mellon) is looking for a job. Any help
or leads you can provide would be greatly appreciated.

[Omar Farook]

Dear Dr. Simon,
  Thanks for information.
  I have two column, and each column contain 4 codes, zero = 0-9, one=10-14, two=15-20, and three=21-30.
  Kindly, the definition of McNemar test talking about two related dichotomous variables, and I don’t have dichotomous variables, my two variables contain 4 categories from 0 to 4 as I have mentioned above , also I don’t have a "before-and-after" data.
  So in this case could I use McNemar test? By the way, I have the original values also.
  Many thanks.
  Omar.


"Simon, Steve, PhD" <[hidden email]> wrote:  Omar writes:

> I have a 72 reading , the first 36 reading are from tooth one
> and the rest from tooth two , the readings are from 36 patient ,
> i.e. two readings per patient. The readings format are 0 , 1 , 2 ,
> and 3 , and each value representing rang, for example , 0
> representing from 0 -9. I want to test if there is a significant
> difference between the two readings of each patient. So by using
> SPSS , is it correct if I use Mann-Whitney U?

I'm not quite sure what the word "rang" means, so perhaps you can
clarify this. But I am quite certain that you do not want the
Mann-Whitney U. This statistic compares two independent groups, and your
groups are clearly related. Two teeth that come from the same mouth
share a common environment and are likely to be strongly correlated. In
fact, you're glad they are strongly correlated, because it improves the
efficiency of your analysis.

If you look at the SPSS menu for nonparametric tests, there is a choice
for "two related samples." This brings up a dialog box which offers the
Wilcoxon, Sign, and McNemar tests. I suspect you want the McNemar test.
Wilcoxon isn't going to work so well because you have a lot of ties in
your data set, and the sign test, though conceptually very simple and
with very few assumptions, just isn't quite as efficient.

If the code 0 represents a range (rang?) of values then you might be
able to gain even more efficiency by using the original values rather
than summarizing them as inside certain ranges. Without more context,
though, I can only speculate.

Finally, if your data is a single column rather than two adjacent
columns, you may need to use the restructure data wizard.

I hope this helps.

Steve Simon, [hidden email], Standard Disclaimer
My niece (Bachelor's degree in Chemical Engineering
from Carnegie Mellon) is looking for a job. Any help
or leads you can provide would be greatly appreciated.



---------------------------------
 Get your own web address.
 Have a HUGE year through Yahoo! Small Business.

[Beadle, ViAnn]

Are the original values scalar? If so, do you still have them? Collapsing to 4 values loses information. If you have the original values, you've got more power to make inferences to a population which is what statistics are all about. Seems to me that it boils down to a simple paired samples t-test.

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Omar Farook
Sent: Tuesday, January 16, 2007 2:14 AM
To: [hidden email]
Subject: Re: is it correct if I use Mann-Whitney U?

Dear Dr. Simon,
  Thanks for information.
  I have two column, and each column contain 4 codes, zero = 0-9, one=10-14, two=15-20, and three=21-30.
  Kindly, the definition of McNemar test talking about two related dichotomous variables, and I don't have dichotomous variables, my two variables contain 4 categories from 0 to 4 as I have mentioned above , also I don't have a "before-and-after" data.
  So in this case could I use McNemar test? By the way, I have the original values also.
  Many thanks.
  Omar.


"Simon, Steve, PhD" <[hidden email]> wrote:  Omar writes:

> I have a 72 reading , the first 36 reading are from tooth one
> and the rest from tooth two , the readings are from 36 patient ,
> i.e. two readings per patient. The readings format are 0 , 1 , 2 ,
> and 3 , and each value representing rang, for example , 0
> representing from 0 -9. I want to test if there is a significant
> difference between the two readings of each patient. So by using
> SPSS , is it correct if I use Mann-Whitney U?

I'm not quite sure what the word "rang" means, so perhaps you can
clarify this. But I am quite certain that you do not want the
Mann-Whitney U. This statistic compares two independent groups, and your
groups are clearly related. Two teeth that come from the same mouth
share a common environment and are likely to be strongly correlated. In
fact, you're glad they are strongly correlated, because it improves the
efficiency of your analysis.

If you look at the SPSS menu for nonparametric tests, there is a choice
for "two related samples." This brings up a dialog box which offers the
Wilcoxon, Sign, and McNemar tests. I suspect you want the McNemar test.
Wilcoxon isn't going to work so well because you have a lot of ties in
your data set, and the sign test, though conceptually very simple and
with very few assumptions, just isn't quite as efficient.

If the code 0 represents a range (rang?) of values then you might be
able to gain even more efficiency by using the original values rather
than summarizing them as inside certain ranges. Without more context,
though, I can only speculate.

Finally, if your data is a single column rather than two adjacent
columns, you may need to use the restructure data wizard.

I hope this helps.

Steve Simon, [hidden email], Standard Disclaimer
My niece (Bachelor's degree in Chemical Engineering
from Carnegie Mellon) is looking for a job. Any help
or leads you can provide would be greatly appreciated.



---------------------------------
 Get your own web address.
 Have a HUGE year through Yahoo! Small Business.

[Simon, Steve, PhD]

Omar Farook writes:

 

> Thanks for information. I have two column, and each column

> contain 4 codes, zero = 0-9, one=10-14, two=15-20, and

> three=21-30. Kindly, the definition of McNemar test talking about two

> related dichotomous variables, and I don't have dichotomous

> variables, my two variables contain 4 categories from 0 to 4 as I have

> mentioned above , also I don't have a "before-and-after" data. So in

> this case could I use McNemar test? By the way, I have the original

> values also. Many thanks. Omar.

 

There are simple extensions of McNemar's test from a dichotomous outcome
variable to a polytomous variable such as yours. These extensions,
unfortunately, are not available in SPSS. Also, while a before and after
design is one example where you might use McNemar's test, any situation
with paired data could be used.

 

If you are willing to suffer the loss of information that occurs when
you collapse a four level category into a two level category, then you
can use SPSS. Otherwise, you have to look elsewhere. A brief Google
search did not yield any software that does this, but I'm sure it is out
there somewhere.

 

You could also analyze the original (uncategorized) values using a
paired t-test.

 

Best of luck with your research.

 

Steve Simon, [hidden email], Standard Disclaimer

My niece (Bachelor's degree in Chemical Engineering

from Carnegie Mellon) is looking for a job. Any help

or leads you can provide would be greatly appreciated.

[Marta García-Granero]

Hi Simon

SSP> There are simple extensions of McNemar's test from a dichotomous outcome
SSP> variable to a polytomous variable such as yours. These extensions,
SSP> unfortunately, are not available in SPSS.

If I remember correctly, SPSS exact tests module includes marginal
homogeneity test, one of the extensions of McNemar test you are
talking about...

Omar, do you have that module (EXACT TESTS) installed in your
computer?

SSP> Also, while a before and after
SSP> design is one example where you might use McNemar's test, any situation
SSP> with paired data could be used.

I'd try Wicoxon test, although I agree that there'll be a lot of tied
data. If sample size is over 25, the asymptotic p-value will be OK and
since it is corrected for ties, there shouldn't be an important loss
of power.


--
Regards,
Dr. Marta García-Granero,PhD           mailto:[hidden email]
Statistician