Is there a way, either within SPSS or outside of it, to conduct nonparametric tests while preserving fractional weights?

[Martha Hewett]

According to the PASW 18 help files, the NPAR tests round weights to the nearest integer.  But the RANK function uses the sum of the case weights, even if they are not integers (I have tested this and found that it used the fractional weight, not a rounded weight).  

Here is my problem:  I have particle events in two different types of environments.  Each event is characterized by two features.  One is its light-absorbing properties for a couple of particular wavelengths (theta), and the other is its size.  I have conducted nonparametric tests of the distribution of events in each environment by theta, and they are different.  That's interesting, but the size of events as a function of theta is also different for the two environments, and both the size and the theta need to be taken into account in looking at difference in people's exposure.  So, I would like to conduct some kind of nonparametric test to compare the distribution of the total mass of events as a function of theta for the two environments.  So I would like to weight by mass (size) and then conduct Kolmogorov-Smirnov tests and perhaps  Mann-Whitney U tests.  The distribution of events by size is probably more or less logarithmic, but the distribution by theta is not any kind of smooth function, so parametric tests are not an option.  

Is there a theoretical reason why you could compute a K-S test or MW-U with integer weights but not non-integer weights?  If not, is there another tool that can do it if SPSS cannot?  (I do have Stata but haven't used it much yet).   Is there a better test for what I'm trying to do?  

Martha Hewett  |

Director of Research | 612.335.5865

Center for Energy and Environment

212 Third Avenue North, Suite 560 | Minneapolis, MN 55401

(cell) 612.839.2358 | (fax) 612.335.5888 | www.mncee.org

[David Marso]

Martha,
Since you do *NOT* describe the specifics of your data (granularity), I can only suggest AGGREGATE the file and then WEIGHT.  You will only be off by n(i) ~.5 .
YMMV, HTH David

Martha Hewett wrote

According to the PASW 18 help files, the NPAR tests round weights to the
nearest integer.  But the RANK function uses the sum of the case weights,
even if they are not integers (I have tested this and found that it used
the fractional weight, not a rounded weight).

Here is my problem:  I have particle events in two different types of
environments.  Each event is characterized by two features.  One is its
light-absorbing properties for a couple of particular wavelengths (theta),
and the other is its size.  I have conducted nonparametric tests of the
distribution of events in each environment by theta, and they are
different.  That's interesting, but the size of events as a function of
theta is also different for the two environments, and both the size and
the theta need to be taken into account in looking at difference in
people's exposure.  So, I would like to conduct some kind of nonparametric
test to compare the distribution of the total mass of events as a function
of theta for the two environments.  So I would like to weight by mass
(size) and then conduct Kolmogorov-Smirnov tests and perhaps  Mann-Whitney
U tests.  The distribution of events by size is probably more or less
logarithmic, but the distribution by theta is not any kind of smooth
function, so parametric tests are not an option.

Is there a theoretical reason why you could compute a K-S test or MW-U
with integer weights but not non-integer weights?  If not, is there
another tool that can do it if SPSS cannot?  (I do have Stata but haven't
used it much yet).   Is there a better test for what I'm trying to do?


Martha Hewett  |
Director of Research | 612.335.5865
Center for Energy and Environment
212 Third Avenue North, Suite 560 | Minneapolis, MN 55401
(cell) 612.839.2358 | (fax) 612.335.5888 | www.mncee.org

[Rich Ulrich]

I can't get past the terminology.  You are writing to social
scientists (for the most part) here, and you might have
some special considerations or conventions that we are
not going to be able to guess.

Is "mass" a reference to particle weight? - If so, I don't
see any justification for counting it as if it were "N" ...
Change the units of weight and you change the N.

But if it is not weight, I don't see where non-integers arise.
And if it is weight, I don't see how it would be justified as a
"weight", counting as multiple observations, for an analysis.


RANK uses fractions, but I would believe instantly that NPAR
rounds off to the nearest integer.  I don't imagine how non-
integers should arrive - legitimately, or usefully -  in your problem,
but that could be my ignorance showing.

It certainly seems that you could be considering transformations
like the frequency (1/theta) for wavelength, in addition to using
the log of weight.  It is no sure thing that parametric tests should
be ruled out because the apparent distribution is not "smooth" --
Does treating the distances as equal (ranks) provide a better
metric that what you get from measurement or a different
transformation than "ranking"?


--
Rich Ulrich

---

Date: Thu, 19 Jan 2012 16:32:14 -0600
From: [hidden email]
Subject: Is there a way, either within SPSS or outside of it, to conduct nonparametric tests while preserving fractional weights?
To: [hidden email]

According to the PASW 18 help files, the NPAR tests round weights to the nearest integer.  But the RANK function uses the sum of the case weights, even if they are not integers (I have tested this and found that it used the fractional weight, not a rounded weight).  

Here is my problem:  I have particle events in two different types of environments.  Each event is characterized by two features.  One is its light-absorbing properties for a couple of particular wavelengths (theta), and the other is its size.  I have conducted nonparametric tests of the distribution of events in each environment by theta, and they are different.  That's interesting, but the size of events as a function of theta is also different for the two environments, and both the size and the theta need to be taken into account in looking at difference in people's exposure.  So, I would like to conduct some kind of nonparametric test to compare the distribution of the total mass of events as a function of theta for the two environments.  So I would like to weight by mass (size) and then conduct Kolmogorov-Smirnov tests and perhaps  Mann-Whitney U tests.  The distribution of events by size is probably more or less logarithmic, but the distribution by theta is not any kind of smooth function, so parametric tests are not an option.  

Is there a theoretical reason why you could compute a K-S test or MW-U with integer weights but not non-integer weights?  If not, is there another tool that can do it if SPSS cannot?  (I do have Stata but haven't used it much yet).   Is there a better test for what I'm trying to do?  

Martha Hewett  |

Director of Research | 612.335.5865

Center for Energy and Environment

212 Third Avenue North, Suite 560 | Minneapolis, MN 55401

(cell) 612.839.2358 | (fax) 612.335.5888 | www.mncee.org

[Art Kendall]

Please explain your study in more detail.

Why are you considering nonparametric approaches?

Is an "event" a case"?  How were they selected?
Please describe your variables,  meaning, level of measurement, value labels, role (IV, DV, something to control), etc.
Are there repeated measures within an event?


On 1/19/2012 5:32 PM, Martha Hewett wrote:

According to the PASW 18 help files, the NPAR tests round weights to the nearest integer.  But the RANK function uses the sum of the case weights, even if they are not integers (I have tested this and found that it used the fractional weight, not a rounded weight).  

Here is my problem:  I have particle events in two different types of environments.  Each event is characterized by two features.  One is its light-absorbing properties for a couple of particular wavelengths (theta), and the other is its size.  I have conducted nonparametric tests of the distribution of events in each environment by theta, and they are different.  That's interesting, but the size of events as a function of theta is also different for the two environments, and both the size and the theta need to be taken into account in looking at difference in people's exposure.  So, I would like to conduct some kind of nonparametric test to compare the distribution of the total mass of events as a function of theta for the two environments.  So I would like to weight by mass (size) and then conduct Kolmogorov-Smirnov tests and perhaps  Mann-Whitney U tests.  The distribution of events by size is probably more or less logarithmic, but the distribution by theta is not any kind of smooth function, so parametric tests are not an option.  

Is there a theoretical reason why you could compute a K-S test or MW-U with integer weights but not non-integer weights?  If not, is there another tool that can do it if SPSS cannot?  (I do have Stata but haven't used it much yet).   Is there a better test for what I'm trying to do?  

Martha Hewett  |

Director of Research | 612.335.5865

Center for Energy and Environment

212 Third Avenue North, Suite 560 | Minneapolis, MN 55401

(cell) 612.839.2358 | (fax) 612.335.5888 | www.mncee.org

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD