| |
 |
|
|
Science Forum Index » Space - Consult Forum » Nonparametric longitudinal analysis...
Page 1 of 1
|
| Author |
Message |
| W... |
 Posted: Fri May 09, 2008 10:42 am |
|
|
|
Guest
|
Hi All:
I have a longitudinal data with 2 factors: starin and age, and a
covariate bodyweight. Each subject was measured at 0min, 30min, 60min,
90min and 120 min. The data were quite right-skewed due to
truncation(the lab machine can only read up to 600unit). I'm planning
to use sas proc mix but I'm afraid the skewness may cause problem in
conclusions. Does anyboday know if there is a nonparametric way to do
this? Is there existing package you would suggest for this type of
analysis?
Thanks! |
|
|
| Back to top |
|
| Richard Ulrich... |
| Posted: Fri May 09, 2008 5:35 pm |
|
|
|
Guest
|
On Fri, 9 May 2008 13:42:35 -0700 (PDT), W <wzhang999 at (no spam) gmail.com>
wrote:
Quote: Hi All:
I have a longitudinal data with 2 factors: starin and age, and a
covariate bodyweight. Each subject was measured at 0min, 30min, 60min,
90min and 120 min.
That sounds simple enough, but I have no idea what
"starin" is (and Google doesn't help). Thus, I have no
natural notion of what is being measured, and whether it
grows/ sinks rapidly, or what.
You can usually get better advice when we readers have
a better idea of what you are trying to accomplish, and
with what data.
Quote: The data were quite right-skewed due to
truncation(the lab machine can only read up to 600unit).
"Right-skewed" in statistics means that there is the
long tail to the right, with the bulk of data on the left.
I think you describe "left-skewed". (This does contradict
a certain natural-language use of skewness, such as,
"attitudes skewed towards Democratic" which would place
the bulk as Democratic.)
Quote: I'm planning
to use sas proc mix but I'm afraid the skewness may cause problem in
conclusions. Does anyboday know if there is a nonparametric way to do
this? Is there existing package you would suggest for this type of
analysis?
How much is "quite right-skewed"?
Are you referring to 10% of the scores, or 60%?
Is this a few persons whose data are all at 600, all
the time (so you might as well toss them out), or do
all the scores tend to hit the ceiling after 60 or 90 minutes?
You don't mention multiple groups. What are you
hoping to show?
For analysis: You might be able to reverse the
scoring (subtract from 600) and have something
that can be represented in Poisson regression, or
Poisson with extra zeroes.
"Nonparametric" -- The traditional choices are actually
somewhat limited. You can rank-order the data by
conventional algorithms (average rank, for ties), and
make your own judgement about whether you like the
scoring that results. Going beyond that step, you could do a
logistic transformation of the ranks.
Otherwise, you could lump all the scores into several
categories, and use ordinal-logistic regression. Or
merely analyze by categories, scored as 1-5 (or so).
--
Rich Ulrich
http://www.pitt.edu/~wpilib/index.html |
|
|
| Back to top |
|
| W... |
| Posted: Mon May 12, 2008 5:13 am |
|
|
|
Guest
|
On May 9, 6:35 pm, Richard Ulrich <Rich.Ulr... at (no spam) comcast.net> wrote:
Quote: On Fri, 9 May 2008 13:42:35 -0700 (PDT), W <wzhang... at (no spam) gmail.com
wrote:
Hi All:
I have a longitudinal data with 2 factors: starin and age, and a
covariate bodyweight. Each subject was measured at 0min, 30min, 60min,
90min and 120 min.
That sounds simple enough, but I have no idea what
"starin" is (and Google doesn't help). Thus, I have no
natural notion of what is being measured, and whether it
grows/ sinks rapidly, or what.
You can usually get better advice when we readers have
a better idea of what you are trying to accomplish, and
with what data.
I apologize. "starin" should be "Strain".
This was an experiment of glucose tolorence test on mice. Glucose was
measured at 0, 30, 60, 90 and 120minutes. The mean at each tome point
was about 177, 537, 516, 512, 508. Time 0 is when glucose was applied.
Quote:
The data were quite right-skewed due to
truncation(the lab machine can only read up to 600unit).
"Right-skewed" in statistics means that there is the
long tail to the right, with the bulk of data on the left.
I think you describe "left-skewed". (This does contradict
a certain natural-language use of skewness, such as,
"attitudes skewed towards Democratic" which would place
the bulk as Democratic.)
Apologize again. I mean it's left-skewed. The response was truncated
at 600 (81 out of 240), which was the ceiling the machine can read.
The Skewness was -0.71 and the Kurtosis was -1.1. Do you think I can
still do the analyses as you proposed above? One thing I was not sure
was how I can handle the correlated structure of the repeated measures
at the 5 time points, which can be easily taken care of by sas proc
mixed if I have a normality assumption for the reponse. |
|
|
| Back to top |
|
| Aniko... |
| Posted: Mon May 12, 2008 5:33 am |
|
|
|
Guest
|
On May 12, 10:13 am, W <wzhang... at (no spam) gmail.com> wrote:
Quote: On May 9, 6:35 pm, Richard Ulrich <Rich.Ulr... at (no spam) comcast.net> wrote:
On Fri, 9 May 2008 13:42:35 -0700 (PDT), W <wzhang... at (no spam) gmail.com
wrote:
Hi All:
I have a longitudinal data with 2 factors: starin and age, and a
covariate bodyweight. Each subject was measured at 0min, 30min, 60min,
90min and 120 min.
That sounds simple enough, but I have no idea what
"starin" is (and Google doesn't help). Thus, I have no
natural notion of what is being measured, and whether it
grows/ sinks rapidly, or what.
You can usually get better advice when we readers have
a better idea of what you are trying to accomplish, and
with what data.
I apologize. "starin" should be "Strain".
This was an experiment of glucose tolorence test on mice. Glucose was
measured at 0, 30, 60, 90 and 120minutes. The mean at each tome point
was about 177, 537, 516, 512, 508. Time 0 is when glucose was applied.
The data were quite right-skewed due to
truncation(the lab machine can only read up to 600unit).
"Right-skewed" in statistics means that there is the
long tail to the right, with the bulk of data on the left.
I think you describe "left-skewed". (This does contradict
a certain natural-language use of skewness, such as,
"attitudes skewed towards Democratic" which would place
the bulk as Democratic.)
Apologize again. I mean it's left-skewed. The response was truncated
at 600 (81 out of 240), which was the ceiling the machine can read.
The Skewness was -0.71 and the Kurtosis was -1.1. Do you think I can
still do the analyses as you proposed above? One thing I was not sure
was how I can handle the correlated structure of the repeated measures
at the 5 time points, which can be easily taken care of by sas proc
mixed if I have a normality assumption for the reponse.- Hide quoted text -
- Show quoted text -
You have bigger problems than just skewness: you have to use a method
that handles the right truncation correctly. For example, your means
are clearly underestimates whenever you have a truncated value. Probit
regression (if the data for a given time point/strain is normal aside
for the truncation issue) and survival analysis have methods for
truncated data. But whatever method you'll choose, it won't be simple.
Aniko |
|
|
| Back to top |
|
| z... |
| Posted: Mon May 12, 2008 10:03 am |
|
|
|
Guest
|
On May 12, 11:13 am, W <wzhang... at (no spam) gmail.com> wrote:
Quote: On May 9, 6:35 pm, Richard Ulrich <Rich.Ulr... at (no spam) comcast.net> wrote:
On Fri, 9 May 2008 13:42:35 -0700 (PDT), W <wzhang... at (no spam) gmail.com
wrote:
Hi All:
I have a longitudinal data with 2 factors: starin and age, and a
covariate bodyweight. Each subject was measured at 0min, 30min, 60min,
90min and 120 min.
That sounds simple enough, but I have no idea what
"starin" is (and Google doesn't help). Thus, I have no
natural notion of what is being measured, and whether it
grows/ sinks rapidly, or what.
You can usually get better advice when we readers have
a better idea of what you are trying to accomplish, and
with what data.
I apologize. "starin" should be "Strain".
This was an experiment of glucose tolorence test on mice. Glucose was
measured at 0, 30, 60, 90 and 120minutes. The mean at each tome point
was about 177, 537, 516, 512, 508. Time 0 is when glucose was applied.
The data were quite right-skewed due to
truncation(the lab machine can only read up to 600unit).
"Right-skewed" in statistics means that there is the
long tail to the right, with the bulk of data on the left.
I think you describe "left-skewed". (This does contradict
a certain natural-language use of skewness, such as,
"attitudes skewed towards Democratic" which would place
the bulk as Democratic.)
Apologize again. I mean it's left-skewed. The response was truncated
at 600 (81 out of 240), which was the ceiling the machine can read.
The Skewness was -0.71 and the Kurtosis was -1.1. Do you think I can
still do the analyses as you proposed above? One thing I was not sure
was how I can handle the correlated structure of the repeated measures
at the 5 time points, which can be easily taken care of by sas proc
mixed if I have a normality assumption for the reponse.- Hide quoted text -
- Show quoted text -
what exactly are you trying to prove/disprove/study?
with data that hits a ceiling like that, you are definitely limited to
certain tests. which i'm trying to remember, having done a lot of this
a while back. Sign test? |
|
|
| Back to top |
|
| Richard Ulrich... |
| Posted: Mon May 12, 2008 11:28 pm |
|
|
|
Guest
|
On Mon, 12 May 2008 08:13:04 -0700 (PDT), W <wzhang999 at (no spam) gmail.com>
wrote:
Quote: On May 9, 6:35 pm, Richard Ulrich <Rich.Ulr... at (no spam) comcast.net> wrote:
On Fri, 9 May 2008 13:42:35 -0700 (PDT), W <wzhang... at (no spam) gmail.com
wrote:
Hi All:
I have a longitudinal data with 2 factors: starin and age, and a
covariate bodyweight. Each subject was measured at 0min, 30min, 60min,
90min and 120 min.
That sounds simple enough, but I have no idea what
"starin" is (and Google doesn't help). Thus, I have no
natural notion of what is being measured, and whether it
grows/ sinks rapidly, or what.
You can usually get better advice when we readers have
a better idea of what you are trying to accomplish, and
with what data.
I apologize. "starin" should be "Strain".
This was an experiment of glucose tolorence test on mice. Glucose was
measured at 0, 30, 60, 90 and 120minutes. The mean at each tome point
was about 177, 537, 516, 512, 508. Time 0 is when glucose was applied.
The data were quite right-skewed due to
truncation(the lab machine can only read up to 600unit).
"Right-skewed" in statistics means that there is the
long tail to the right, with the bulk of data on the left.
I think you describe "left-skewed". (This does contradict
a certain natural-language use of skewness, such as,
"attitudes skewed towards Democratic" which would place
the bulk as Democratic.)
Apologize again. I mean it's left-skewed. The response was truncated
at 600 (81 out of 240), which was the ceiling the machine can read.
The Skewness was -0.71 and the Kurtosis was -1.1. Do you think I can
still do the analyses as you proposed above? One thing I was not sure
was how I can handle the correlated structure of the repeated measures
at the 5 time points, which can be easily taken care of by sas proc
mixed if I have a normality assumption for the reponse.
Okay. One thing that seems *apparent* from the means
presented is that the 0-time is low, and the highest point is
the first measurement after that. If that was not a stupid and
fatal truncation of the data, then the "effect" you are investigating
would be essentially uninterpretable from the general repeated
measures analysis --
1) There is a difference between times.
2) There are linear, quadratic, cubic, and quartic terms that
are significant. Or some combination of most of them.
So what?
What seems to be interesting, if anything, is that there may
be a decline after the initial boost. For that, you use an
analysis of the latter four times. Look at the linear trend.
That test is pretty robust.
I suspect that the heterogeneity of variance will not be
bad for these 4 times, and there is the conventional correction
available for the overall repeated-measures F-test d.f. (if you
want to look at that test), since 80 or so measures (out of 192)
are truncated at 600.
--
Rich Ulrich
http://www.pitt.edu/~wpilib/index.html |
|
|
| Back to top |
|
| |
|
Page 1 of 1
All times are GMT - 5 Hours
The time now is Thu Jul 24, 2008 11:16 pm
|
|
