SPSSX Discussion

Unequal size groups

Classic

List

Threaded

6 messages Options

Triantafillos Pliakas

Unequal size groups

Dear List,

I am doing analysis on some data I have on children
aged from birth to 10 years. I have classified
children by age group (in one year intervals), gender
and also by the presence of obesity (Initially I had
three groups -normal weight, overweight and obese- but
due to the small number of overweight and obese
children I grouped them all together in order to
acquire more statistical power; So now I have two
groups - normal weight and overweight-).

I want to assess the differences in a continuous
variable (waist to height ratio) between normal weight
and overweight children, by gender and age group. I
have explored the data using the Kolmogorov-Smirnof
test and Shapiro-Wilk test, excluded potential
outliers and found that in most cases (except only one
age group) the above tests were significant. So
instead of using a parametric test (independent t
test) or even try to transform the data
logarithmically I decided to use the non parapetric
test Mann-Whitney. My questions are:

1) Can I use the Mann-Whitney test even though my test
variable (waist to height ratio) is not ordinal?

2) Even if I can use Mann-Whitney test (or even if I
need to carry out another test) the thing is that I
still have a very small number of overweight children
(even after grouping overweight and obese children).
The number of normal weight children range from 226 to
398 whereas for overweight range from 13 to 44 between
the different age groups(For example I have 320 normal
weight boys and 34 overweight ones in age group 6-7
years and in age group 8-9 years I have 398 normal
weight girls and 13 overweight ones). Does this
underestimate or distort any statistical differences
found? (I assume it does) And if so is there a better
statistical test to carry out to minimize this or at
least to address this more appropriately?

Thank you in advance

Triantafyllos

___________________________________________________________
Χρησιμοποιείτε Yahoo!;
Βαρεθήκατε τα ενοχλητικά μηνύματα (spam); Το Yahoo! Mail
διαθέτει την καλύτερη δυνατή προστασία κατά των ενοχλητικών
μηνυμάτων http://login.yahoo.com/config/mail?.intl=gr

Dominic Lusinchi

Re: Unequal size groups

Dear colleague,

A general remark: it is not a good idea to exclude "outliers" - unless you
have some very solid reasons to do so.

You can use the MW test for any variable that is measured at least at the
ordinal level - so you can certainly use it for an interval or ratio level
variable.

You can use the MW test when the two groups are of unequal size.

However, as I understand you situation, you have more than one independent
variable, or perhaps covariate. It would seem that a factorial approach
would be more appropriate: the ratio is the dependent variable, the two
weight groups the independent variable, and age and gender as the covariates
(?).

Good luck.

Dominic Lusinchi
Statistician
Far West Research
Statistical Consulting
San Francisco, California
415-664-3032
www.farwestresearch.com

-----Original Message-----
From: Triantafillos Pliakas [mailto:[hidden email]]
Sent: Friday, February 16, 2007 2:40 PM
Subject: Unequal size groups

Dear List,

I am doing analysis on some data I have on children
aged from birth to 10 years. I have classified
children by age group (in one year intervals), gender
and also by the presence of obesity (Initially I had
three groups -normal weight, overweight and obese- but
due to the small number of overweight and obese
children I grouped them all together in order to
acquire more statistical power; So now I have two
groups - normal weight and overweight-).

I want to assess the differences in a continuous
variable (waist to height ratio) between normal weight
and overweight children, by gender and age group. I
have explored the data using the Kolmogorov-Smirnof
test and Shapiro-Wilk test, excluded potential
outliers and found that in most cases (except only one
age group) the above tests were significant. So
instead of using a parametric test (independent t
test) or even try to transform the data
logarithmically I decided to use the non parapetric
test Mann-Whitney. My questions are:

1) Can I use the Mann-Whitney test even though my test
variable (waist to height ratio) is not ordinal?

2) Even if I can use Mann-Whitney test (or even if I
need to carry out another test) the thing is that I
still have a very small number of overweight children
(even after grouping overweight and obese children).
The number of normal weight children range from 226 to
398 whereas for overweight range from 13 to 44 between
the different age groups(For example I have 320 normal
weight boys and 34 overweight ones in age group 6-7
years and in age group 8-9 years I have 398 normal
weight girls and 13 overweight ones). Does this
underestimate or distort any statistical differences
found? (I assume it does) And if so is there a better
statistical test to carry out to minimize this or at
least to address this more appropriately?

Thank you in advance

Triantafyllos

___________________________________________________________
Χρησιμοποιείτε Yahoo!;
Βαρεθήκατε τα ενοχλητικά μηνύματα (spam); Το Yahoo! Mail
διαθέτει την καλύτερη δυνατή προστασία κατά των ενοχλητικών
μηνυμάτων http://login.yahoo.com/config/mail?.intl=gr

Dominic Lusinchi

RE: Θέμα: RE: Unequal size groups

Dear Triantafylos,

I answer you on the list because that way folks on the list can follow the
thread and make any contributions, corrections, comments as they see fit.

You do not say how the distributions on the dependent differ: "not the same"
does not tell us much. But to answer your question: when comparing two
groups with the MW test, the distribution of the dependent in the two groups
should be of a similar shape: both positively skewed, for example.

Good luck.

Dominic Lusinchi
Statistician
Far West Research
Statistical Consulting
San Francisco, California
415-664-3032
www.farwestresearch.com

-----Original Message-----
From: Triantafillos Pliakas [mailto:[hidden email]]
Sent: Saturday, February 17, 2007 4:23 PM
To: Dominic Lusinchi
Subject: Θέμα: RE: Unequal size groups

Dear Dominic,

thanks for clarifying things with MW test. Another
issue that is raised here is what happens if the
distributions of the dependent variable (in this case
the ratio) are not the same in the two groups (normal
and overweight)? Using the two sample KS test I found
out that in some age groups the distributions are not
the same. I presume these raises an issue as to the
appropriateness of using MW test after all.

Regarding outliers of course I agree with you but in
my analysis it was very clear that some of the cases
were outliers (such cases were children with a BMI z
score -26.75 or waist circumference z score -12.41 -
the population under study was a healthy one and
therefore values such as the above were considered to
be outliers, probably a mistake was made while
entering the data but I didnt want to alter any of
those values-).

I will try factorial analysis as you have suggested
but my statistical background is quite limited so it
will take me sometime to figure things out.

Thanks again for your help.

Triantafyllos

--- Dominic Lusinchi <[hidden email]>
έγραψε:

> Dear colleague,
>
> A general remark: it is not a good idea to exclude
> "outliers" - unless you
> have some very solid reasons to do so.
>
> You can use the MW test for any variable that is
> measured at least at the
> ordinal level - so you can certainly use it for an
> interval or ratio level
> variable.
>
> You can use the MW test when the two groups are of
> unequal size.
>
> However, as I understand you situation, you have
> more than one independent
> variable, or perhaps covariate. It would seem that a
> factorial approach
> would be more appropriate: the ratio is the
> dependent variable, the two
> weight groups the independent variable, and age and
> gender as the covariates
> (?).
>
> Good luck.
>
> Dominic Lusinchi
> Statistician
> Far West Research
> Statistical Consulting
> San Francisco, California
> 415-664-3032
> www.farwestresearch.com
>
> -----Original Message-----
> From: Triantafillos Pliakas
> [mailto:[hidden email]]
> Sent: Friday, February 16, 2007 2:40 PM
> Subject: Unequal size groups
>
> Dear List,
>
> I am doing analysis on some data I have on children
> aged from birth to 10 years. I have classified
> children by age group (in one year intervals),
> gender
> and also by the presence of obesity (Initially I had
> three groups -normal weight, overweight and obese-
> but
> due to the small number of overweight and obese
> children I grouped them all together in order to
> acquire more statistical power; So now I have two
> groups - normal weight and overweight-).
>
> I want to assess the differences in a continuous
> variable (waist to height ratio) between normal
> weight
> and overweight children, by gender and age group. I
> have explored the data using the Kolmogorov-Smirnof
> test and Shapiro-Wilk test, excluded potential
> outliers and found that in most cases (except only
> one
> age group) the above tests were significant. So
> instead of using a parametric test (independent t
> test) or even try to transform the data
> logarithmically I decided to use the non parapetric
> test Mann-Whitney. My questions are:
>
> 1) Can I use the Mann-Whitney test even though my
> test
> variable (waist to height ratio) is not ordinal?
>
> 2) Even if I can use Mann-Whitney test (or even if I
> need to carry out another test) the thing is that I
> still have a very small number of overweight
> children
> (even after grouping overweight and obese children).
> The number of normal weight children range from 226
> to
> 398 whereas for overweight range from 13 to 44
> between
> the different age groups(For example I have 320
> normal
> weight boys and 34 overweight ones in age group 6-7
> years and in age group 8-9 years I have 398 normal
> weight girls and 13 overweight ones). Does this
> underestimate or distort any statistical differences
> found? (I assume it does) And if so is there a
> better
> statistical test to carry out to minimize this or at
> least to address this more appropriately?
>
> Thank you in advance
>
> Triantafyllos
>
>
>
>
>
>
>
>
>

___________________________________________________________

> Χρησιμοποιείτε Yahoo!;
> Βαρεθήκατε τα ενοχλητικά μηνύματα (spam); Το Yahoo!
> Mail
> διαθέτει την καλύτερη δυνατή προστασία κατά των
> ενοχλητικών
> μηνυμάτων
> http://login.yahoo.com/config/mail?.intl=gr
>
>
>

___________________________________________________________
Χρησιμοποιείτε Yahoo!;
Βαρεθήκατε τα ενοχλητικά μηνύματα (spam); Το Yahoo! Mail
διαθέτει την καλύτερη δυνατή προστασία κατά των ενοχλητικών
μηνυμάτων http://login.yahoo.com/config/mail?.intl=gr

Triantafillos Pliakas

Θέμα: RE: Θέμα: RE: Unequal size groups

My apologies to the list and to you.

I assumed that when the two sample Kolmogorov-Smirnov
test is significant then this shows that the
distribution of the dependant variable (the ratio) is
not the same (either in shape or location) in the two
groups (normal and overweight).

I didn't examine how those distributions differ. But
after you comment I did so by plotting the data and by
getting the value for skewness. In some age groups the
distribution did not have the same shape (were either
positevely or negatively skewed in one group or the
other). I understand that this consists a violation of
the assumption in terms of the distribution and
therefore MW test can not be used. Is that correct?

Regards
Triantafyllos

--- Dominic Lusinchi <[hidden email]>
έγραψε:

> Dear Triantafylos,
>
> I answer you on the list because that way folks on
> the list can follow the
> thread and make any contributions, corrections,
> comments as they see fit.
>
> You do not say how the distributions on the
> dependent differ: "not the same"
> does not tell us much. But to answer your question:
> when comparing two
> groups with the MW test, the distribution of the
> dependent in the two groups
> should be of a similar shape: both positively
> skewed, for example.
>
> Good luck.
>
> Dominic Lusinchi
> Statistician
> Far West Research
> Statistical Consulting
> San Francisco, California
> 415-664-3032
> www.farwestresearch.com
>
> -----Original Message-----
> From: Triantafillos Pliakas
> [mailto:[hidden email]]
> Sent: Saturday, February 17, 2007 4:23 PM
> To: Dominic Lusinchi
> Subject: Θέμα: RE: Unequal size groups
>
> Dear Dominic,
>
> thanks for clarifying things with MW test. Another
> issue that is raised here is what happens if the
> distributions of the dependent variable (in this
> case
> the ratio) are not the same in the two groups
> (normal
> and overweight)? Using the two sample KS test I
> found
> out that in some age groups the distributions are
> not
> the same. I presume these raises an issue as to the
> appropriateness of using MW test after all.
>
> Regarding outliers of course I agree with you but in
> my analysis it was very clear that some of the cases
> were outliers (such cases were children with a BMI z
> score -26.75 or waist circumference z score -12.41 -
> the population under study was a healthy one and
> therefore values such as the above were considered
> to
> be outliers, probably a mistake was made while
> entering the data but I didnt want to alter any of
> those values-).
>
> I will try factorial analysis as you have suggested
> but my statistical background is quite limited so it
> will take me sometime to figure things out.
>
> Thanks again for your help.
>
> Triantafyllos
>
> --- Dominic Lusinchi <[hidden email]>
> έγραψε:
>
> > Dear colleague,
> >
> > A general remark: it is not a good idea to exclude
> > "outliers" - unless you
> > have some very solid reasons to do so.
> >
> > You can use the MW test for any variable that is
> > measured at least at the
> > ordinal level - so you can certainly use it for an
> > interval or ratio level
> > variable.
> >
> > You can use the MW test when the two groups are of
> > unequal size.
> >
> > However, as I understand you situation, you have
> > more than one independent
> > variable, or perhaps covariate. It would seem that
> a
> > factorial approach
> > would be more appropriate: the ratio is the
> > dependent variable, the two
> > weight groups the independent variable, and age
> and
> > gender as the covariates
> > (?).
> >
> > Good luck.
> >
> > Dominic Lusinchi
> > Statistician
> > Far West Research
> > Statistical Consulting
> > San Francisco, California
> > 415-664-3032
> > www.farwestresearch.com
> >
> > -----Original Message-----
> > From: Triantafillos Pliakas
> > [mailto:[hidden email]]
> > Sent: Friday, February 16, 2007 2:40 PM
> > Subject: Unequal size groups
> >
> > Dear List,
> >
> > I am doing analysis on some data I have on
> children
> > aged from birth to 10 years. I have classified
> > children by age group (in one year intervals),
> > gender
> > and also by the presence of obesity (Initially I
> had
> > three groups -normal weight, overweight and obese-
> > but
> > due to the small number of overweight and obese
> > children I grouped them all together in order to
> > acquire more statistical power; So now I have two
> > groups - normal weight and overweight-).
> >
> > I want to assess the differences in a continuous
> > variable (waist to height ratio) between normal
> > weight
> > and overweight children, by gender and age group.
> I
> > have explored the data using the
> Kolmogorov-Smirnof
> > test and Shapiro-Wilk test, excluded potential
> > outliers and found that in most cases (except only
> > one
> > age group) the above tests were significant. So
> > instead of using a parametric test (independent t
> > test) or even try to transform the data
> > logarithmically I decided to use the non
> parapetric
> > test Mann-Whitney. My questions are:
> >
> > 1) Can I use the Mann-Whitney test even though my
> > test
> > variable (waist to height ratio) is not ordinal?
> >
> > 2) Even if I can use Mann-Whitney test (or even if
> I
> > need to carry out another test) the thing is that
> I
> > still have a very small number of overweight
> > children
> > (even after grouping overweight and obese
> children).
> > The number of normal weight children range from
> 226
> > to
> > 398 whereas for overweight range from 13 to 44
> > between
> > the different age groups(For example I have 320
> > normal
> > weight boys and 34 overweight ones in age group
> 6-7
> > years and in age group 8-9 years I have 398 normal
> > weight girls and 13 overweight ones). Does this
> > underestimate or distort any statistical
> differences
> > found? (I assume it does) And if so is there a
> > better
> > statistical test to carry out to minimize this or
> at
> > least to address this more appropriately?
> >
> > Thank you in advance
> >
> > Triantafyllos
> >
> >
> >
> >
> >
> >
> >
> >
> >
>

___________________________________________________________

> > Χρησιμοποιείτε Yahoo!;
> > Βαρεθήκατε τα ενοχλητικά μηνύματα (spam); Το
> Yahoo!
> > Mail
> > διαθέτει την καλύτερη δυνατή προστασία κατά των
> > ενοχλητικών
> > μηνυμάτων
> > http://login.yahoo.com/config/mail?.intl=gr
> >
> >
> >
>

=== message truncated ===

___________________________________________________________
Χρησιμοποιείτε Yahoo!;
Βαρεθήκατε τα ενοχλητικά μηνύματα (spam); Το Yahoo! Mail
διαθέτει την καλύτερη δυνατή προστασία κατά των ενοχλητικών
μηνυμάτων http://login.yahoo.com/config/mail?.intl=gr

Dominic Lusinchi

RE: Θέμα: RE: Θέμα: RE: Unequal size groups

No need for apologies.

It is always a good idea (as I'm sure you already know) to use
Analyze>Descriptive Statistics>Explore... when you are working on any data
set. With that procedure you get not only useful statistics but a number of
graphs that allow you to visualize your data. That way, after you've
conducted a formal test, you can assess if it confirms what you have
discovered visually with the graphs, and numerically with the statistics.

If you have used the K-S test why do you want to use the M-W test also?

In any case, I would suggest that you take a look at the examples provided
by SPSS under the help menu on both these tests. I think you will find very
useful for your purpose.

Good luck.

Dominic Lusinchi
Statistician
Far West Research
Statistical Consulting
San Francisco, California
415-664-3032
www.farwestresearch.com

-----Original Message-----
From: Triantafillos Pliakas [mailto:[hidden email]]
Sent: Sunday, February 18, 2007 4:54 PM
To: [hidden email]
Cc: Dominic Lusinchi
Subject: Θέμα: RE: Θέμα: RE: Unequal size groups

My apologies to the list and to you.

I assumed that when the two sample Kolmogorov-Smirnov
test is significant then this shows that the
distribution of the dependant variable (the ratio) is
not the same (either in shape or location) in the two
groups (normal and overweight).

I didn't examine how those distributions differ. But
after you comment I did so by plotting the data and by
getting the value for skewness. In some age groups the
distribution did not have the same shape (were either
positevely or negatively skewed in one group or the
other). I understand that this consists a violation of
the assumption in terms of the distribution and
therefore MW test can not be used. Is that correct?

Regards
Triantafyllos

--- Dominic Lusinchi <[hidden email]>
έγραψε:

___________________________________________________________

Triantafillos Pliakas

Θέμα: RE: Θέμα: RE: Θέμα: RE: Unequal size group s

The first thing I always do is to explore the data as
you have commented (Analyze>Descriptive
Statistics>Explore).

I used the KS test to see if the distributions of the
ratio in the two groups are of similar shape as this
is a prerequisite for using MW test. I understand why
you questioned the use of one test instead of the
other. I knew that the non parametric test for
assessing differences between groups is the MW test
and then I discovered that in order to use MW test the
distributions must have the same shape. But apparently
KS can also be used for assessing differences between
two groups which I didnt know until now.

Thanks and I will have a look at the factorial
analysis as you have suggested.

Triantafyllos

--- Dominic Lusinchi <[hidden email]>
έγραψε:

> No need for apologies.
>
> It is always a good idea (as I'm sure you already
> know) to use
> Analyze>Descriptive Statistics>Explore... when you
> are working on any data
> set. With that procedure you get not only useful
> statistics but a number of
> graphs that allow you to visualize your data. That
> way, after you've
> conducted a formal test, you can assess if it
> confirms what you have
> discovered visually with the graphs, and numerically
> with the statistics.
>
> If you have used the K-S test why do you want to use
> the M-W test also?
>
> In any case, I would suggest that you take a look at
> the examples provided
> by SPSS under the help menu on both these tests. I
> think you will find very
> useful for your purpose.
>
> Good luck.
>
> Dominic Lusinchi
> Statistician
> Far West Research
> Statistical Consulting
> San Francisco, California
> 415-664-3032
> www.farwestresearch.com
>
> -----Original Message-----
> From: Triantafillos Pliakas
> [mailto:[hidden email]]
> Sent: Sunday, February 18, 2007 4:54 PM
> To: [hidden email]
> Cc: Dominic Lusinchi
> Subject: Θέμα: RE: Θέμα: RE: Unequal size groups
>
> My apologies to the list and to you.
>
> I assumed that when the two sample
> Kolmogorov-Smirnov
> test is significant then this shows that the
> distribution of the dependant variable (the ratio)
> is
> not the same (either in shape or location) in the
> two
> groups (normal and overweight).
>
> I didn't examine how those distributions differ. But
> after you comment I did so by plotting the data and
> by
> getting the value for skewness. In some age groups
> the
> distribution did not have the same shape (were
> either
> positevely or negatively skewed in one group or the
> other). I understand that this consists a violation
> of
> the assumption in terms of the distribution and
> therefore MW test can not be used. Is that correct?
>
>
> Regards
> Triantafyllos
>
>
>
> --- Dominic Lusinchi <[hidden email]>
> έγραψε:
>
> > Dear Triantafylos,
> >
> > I answer you on the list because that way folks on
> > the list can follow the
> > thread and make any contributions, corrections,
> > comments as they see fit.
> >
> > You do not say how the distributions on the
> > dependent differ: "not the same"
> > does not tell us much. But to answer your
> question:
> > when comparing two
> > groups with the MW test, the distribution of the
> > dependent in the two groups
> > should be of a similar shape: both positively
> > skewed, for example.
> >
> > Good luck.
> >
> > Dominic Lusinchi
> > Statistician
> > Far West Research
> > Statistical Consulting
> > San Francisco, California
> > 415-664-3032
> > www.farwestresearch.com
> >
> > -----Original Message-----
> > From: Triantafillos Pliakas
> > [mailto:[hidden email]]
> > Sent: Saturday, February 17, 2007 4:23 PM
> > To: Dominic Lusinchi
> > Subject: Θέμα: RE: Unequal size groups
> >
> > Dear Dominic,
> >
> > thanks for clarifying things with MW test. Another
> > issue that is raised here is what happens if the
> > distributions of the dependent variable (in this
> > case
> > the ratio) are not the same in the two groups
> > (normal
> > and overweight)? Using the two sample KS test I
> > found
> > out that in some age groups the distributions are
> > not
> > the same. I presume these raises an issue as to
> the
> > appropriateness of using MW test after all.
> >
> > Regarding outliers of course I agree with you but
> in
> > my analysis it was very clear that some of the
> cases
> > were outliers (such cases were children with a BMI
> z
> > score -26.75 or waist circumference z score -12.41
> -
> > the population under study was a healthy one and
> > therefore values such as the above were considered
> > to
> > be outliers, probably a mistake was made while
> > entering the data but I didnt want to alter any of
> > those values-).
> >
> > I will try factorial analysis as you have
> suggested
> > but my statistical background is quite limited so
> it
> > will take me sometime to figure things out.
> >
> > Thanks again for your help.
> >
> > Triantafyllos
> >
> > --- Dominic Lusinchi <[hidden email]>
> > έγραψε:
> >
> > > Dear colleague,
> > >
> > > A general remark: it is not a good idea to
> exclude
> > > "outliers" - unless you
> > > have some very solid reasons to do so.
> > >
> > > You can use the MW test for any variable that is
> > > measured at least at the
> > > ordinal level - so you can certainly use it for
> an
> > > interval or ratio level
> > > variable.
> > >
> > > You can use the MW test when the two groups are
> of
> > > unequal size.
> > >
> > > However, as I understand you situation, you have
> > > more than one independent
> > > variable, or perhaps covariate. It would seem
> that
> > a
> > > factorial approach
> > > would be more appropriate: the ratio is the
> > > dependent variable, the two
> > > weight groups the independent variable, and age
> > and
> > > gender as the covariates
> > > (?).
> > >
> > > Good luck.
> > >
> > > Dominic Lusinchi
> > > Statistician
> > > Far West Research
> > > Statistical Consulting
> > > San Francisco, California
> > > 415-664-3032
> > > www.farwestresearch.com
> > >
> > > -----Original Message-----
> > > From: Triantafillos Pliakas
> > > [mailto:[hidden email]]
> > > Sent: Friday, February 16, 2007 2:40 PM
> > > Subject: Unequal size groups
> > >
> > > Dear List,
>