Subject: Re: Experimental Design Puzzle
From: rbcrosie@apgea.army.mil (Ronald B. Crosier)
Date: Fri, 17 Jan 97 17:09:04 GMT
My understanding is that you have 480 measurements of soil stability,
and you want to know how (or whether) these measurements depend on soil
type/location and aggregate size, but you realize that they may also
depend on the soil sample. So, in addition to the random measurement
error, there is another random error associated with soil sample. In
my humble opinion, you have a three-way analysis of variance (ANOVA):
Factor 1: soil type/location, 8 levels
Factor 2: sample, 10 levels, nested within soil type/location
Factor 3: aggregate size, 6 levels
Factors 1 and 3 are fixed effects; Factor 2 is a random effect.
I believe that treating factor 2 as a random, nested effect correctly
accounts for what you call the "repeated measures" nature of the design.
Here is my guess for the ANOVA terms, their degrees of freedom and
error terms.
Term df Error Term
grand mean 1
soil 7 sample(soil)
sample(soil) 72 measurement error
size 5 size*sample(soil)
soil*size 35 size*sample(soil)
size*sample(soil) 360 measurement error
measurement error 0
total 480
--
Ronald Crosier E-mail:
Disclaimer: My opinions are just that---mine, and opinions.
Subject: Re: Simpson's paradox
From: Clay Helberg
Date: Fri, 17 Jan 1997 11:29:27 -0600
John R. Vokey wrote:
> Anil Menon (Syracuse University, School of Engg. and Computer Science)
> compiled this reference list and placed it on a "Simpson's Paradox" web-site
> he had created. Unfortunately, the web-site URL has changed (or no longer
> exists):
>
Wow, what a great resource! Thanks for sharing it. You might also want
to add the paper from the latest issue of American Statistician to the
list:
Appleton, Frnech, & Vanderpump (1996). Ignoring a covariate: an example
of Simpson's paradox. American Statistician, 50(4), 340-341.
--
Clay Helberg | Internet: helberg@execpc.com
Publications Dept. | WWW: http://www.execpc.com/~helberg/
SPSS, Inc. | Speaking only for myself....
Subject: Simpson's paradox
From: Scott Chasalow
Date: Fri, 17 Jan 1997 10:19:40 GMT
Yet one more reference on Simpson's paradox (and related phenomena),
on the theoretical side:
Samuels, Myra L (1993) Simpson's Paradox and Related Phenomena. Journal of
the American Statistical Association 88: 81-88.
She talks about "association reversal", a sort of generalization of Simpson's
paradox. I've found her results useful at least once, in explaining a
spurious association seen in genetic mapping of a quantitative trait,
which entails looking for associations between discrete marker genotypes and
continuous trait values, rather than among discrete variables only.
Scott
----------------------------------
Scott D. Chasalow
Department of Plant Breeding
Wageningen Agricultural University
Scott.Chasalow@users.pv.wau.nl
----------------------------------
Subject: Re: Mokken Scales___MSP
From: Jean-Luc KOP
Date: Fri, 17 Jan 1997 08:57:22 -0100
check the following url http://www.progamma.nl
hope this helps
Jean-Luc
At 19:06 16/01/1997 -0000, you wrote:
>Hello,
>
>I have a copy of MSP, a program for Mokken Scale
>developed by DISC, University of Amsterdam.
>
>I have an old copy version 1.50 (1988) and
>I have lost the connection with them and they
>send me nothing in all this years.
>
>Please, can anyone give some address to contact
>about MSP and in general about software for IRT
>
>Thanks in advance
>Jose
>
>
ooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooo
oo Jean-Luc KOP oo
oo oo
oo GRAPCO ADEPS oo
oo Laboratoire de Psychologie URA CNRS 1167 oo
oo Universite Nancy 2 Universite Nancy 2 oo
oo B.P. 3397 C.O. 26 oo
oo 54015 Nancy Cedex FRANCE 54035 Nancy Cedex FRANCE oo
oo oo
oo Tel : (33) 03 83 96 71 90 Tel : (33) 03 83 30 58 41 oo
oo Fax : (33) 03 83 96 70 90 Fax : (33) 03 83 35 83 92 oo
oo oo
oo e-mail : kop@clsh.u-nancy.fr oo
ooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooo
Subject: Re: Joint confidence interval-degrees of freedom
From: wpilib+@pitt.edu (Richard F Ulrich)
Date: 17 Jan 1997 18:59:24 GMT
Jordi Riu (rusell@QUIMICA.URV.ES) wrote:
: Hello,
: Let's suppose I have a multivariate model, y=b0+b1x1+b2x2. I can buid the
: joint confidence interval for the b0, b1 and b2 coefficients as the joint
: confidence ellipsoid, which follows an F test with alpha level and 3 and n-3
: degrees of freedom. Now I take the former model, and I combine b1 and b2 in
: the following way: c=b1+b2 in order to rewrite the model: y=b0+cx'. My
: problem concerns to the degrees of freedom of the joint confidence interval
: of the new model. Are they 2 and n-2 or 3 and n-3 (or something else)?
-- When I have one variable, I often think of the interval around
the b as being the simple t distance - instead of including the
intercept, for a model with 2 d.f. Does anybody have a comment on
when, or why, 1 vs. 2 d.f. should be used with one variable?
I do have a comment about your two versions - first, what is
written is typographically wrong or unclear. Whatever might it
mean, to combine "c=b1+b2"? - where b1 and b2 were coefficients for
two different variables.
If you are referring to entering x1 and x2 separately, as opposed to
a model using the sum of (x1+x2):
-- If there were good reason to consider the (x1+x2)
model before you looked at your data, then, Sure, you can consider
the sum to be one variable which only accounts for one d.f. in the
statistics. It is a simple composite variable.
-- If you only decided to look at a sum *because* the other model
showed b1, b2 to be nearly equal, then, No, if the loss of a
single d.f. hurts you since you really have thin evidence for your
model, you better accept the loss of the d.f.
Rich Ulrich, biostatistician wpilib+@pitt.edu
http://www.pitt.edu/~wpilib/index.html Univ. of Pittsburgh
Subject: Re: Percentage error in fit
From: wpilib+@pitt.edu (Richard F Ulrich)
Date: 17 Jan 1997 19:19:14 GMT
<< S.Kannan >>
Prof. K. Kishore (kishore@HAMSADVANI.SERC.IISC.ERNET.IN) wrote:
: Hi listmembers,
: I am a novice in statistics. I would like to link
: kinematic viscosity (v) to time taken (t) for a liquid to flow
: between two marks by a relationship
: v = at - b/t
: where a and b are fitting parameters. Now my question is how
: to calculate the error in the fit by this function. I have done
: this way
: s = sqrt(sum((Y rep - Y calc)^2)/N)
: where N is the number of data points. Can I report this
: parameter (s) as the error in the fit ?. Is it possible to find
: percentage error in this case ?.
You are computing the "root mean-square()" for something, which
is one reasonable way to start something, which could be part of
an error report. Are "Y rep" and "Y calc" the values
for v, in the data, and then as estimated from t ? You could
report the error in this fit, as compared to the error from just
fitting the MEAN of v - that is what you are given with regression
calculations, so you would be talking about some conventional
quantities.
Rich Ulrich, biostatistician wpilib+@pitt.edu
http://www.pitt.edu/~wpilib/index.html Univ. of Pittsburgh
Subject: Re: kappa
From: Alan Hutson
Date: Fri, 17 Jan 1997 08:08:25 -0500
Anyone interested in a comparitive study of measuring agreement
between raters should check out
Bartko (1994) Measures of Agreement: A Single Procedure. Statistics in
Medicine (13) 737-745.
He gives illustrations where the ICC, Pearson's correlation, the t-test
etc. can all be misleading measures of agreement. The recommended method
of assessing agreement is more graphical Bland-Altmann type stuff.
Alan
>
> as you have clearly pointed out, I am obviously a dunce about this;
> and thus:
>
> I still wonder about the case were two raters may have the same mean yet poor
> ICC coefficient and I wonder about the case where
> raters 1, 2, and 3, could have means that were in that order, that is,
> mean 2 could be closer to mean 1 than mean 3 but:
> means 1 and 3 could have better reliability (w/ ICC) than means 1 and 2;
>
> I still am not sure why you would rather do a t-test and a Pearson's C
>
> Since in calculating an ICC you would have the results of an ANOVA to test "
> if one rater or rating is systematically higher than
> another." and you would have a correlation coefficient that as I understand
> the liturature to indicate contains more information than Pearson's since the
> ICC incorporates association and agreement rather than only association.
>
>
>
> Chauncey wrote:
> > : snip . . .
> >
> > : If your measure is interval like, ICC is the interrater reliabability
> > : stat to use; I would think. but alas, I'm still quite a naive student.
>
> Richard F Ulrich wrote:
> >
> > There seems to me that there must be some blindness in the way that
> > "reliability" is being taught, because the point that I was making is
> > a simple one... yet, this is not the first time that it has been
> > missed.
> >
> > The intra-class correlation (ICC) is a fine measurement for
> > publishing what you have achieved in "reliability". Unfortunately,
> > it does nothing to illustrate or test or separate out the
> > *systematic differences* that may occur between raters - they
> > just serve to lower the correlation slightly, since the ICC makes
> > the assumption that the raters have equal means.
> >
> > In almost any kind of work that I can think of, it *ought* to be a
> > concern if one rater or rating is systematically higher than
> > another. The powerful way to test this is with the paired t-test;
> > the concommitant statistic to the paired t-test is the Pearson
> > correlation - together they give both aspects of comparing the
> > ratings, SIMILARITY and DIFFERENCE.
> >
> > So, the ICC may be what editors want to see, and it is okay as
> > a one-number summary, but anyone examining their own reliability
> > data has little excuse (IMHO) not to look at tests of difference,
> > where they are appropriate.
> >
> > Rich Ulrich, biostatistician wpilib+@pitt.edu
> > http://www.pitt.edu/~wpilib/index.html Univ. of Pittsburgh
>
Subject: Re: Negative confidence interval for proportion ????
From: Warren
Date: 17 Jan 1997 21:16:27 GMT
j_weedon@escape.com (Jay Weedon) wrote:
>On 15 Jan 1997 22:32:41 GMT, tseck@gibbs.oit.unc.edu (Chiu Kit Jessica
>Tse) wrote:
>
>>I calculated the weighted proportion and standard error
>>of a rare factor.
>>
>>I got the proportion as 0.0175 and
>> the standard error as 0.011895.
>>
>Exact confidence limits for proportions can be found in many intro
>stats texts, e.g., Zar JH (1984) Biostatistical Analysis. Englewood
>Cliffs NJ: Prentice-Hall. (2nd ed) p.378.
>
In an earlier post, I mentioned using Quesenberry and Hurst's ideas as
mentioned by Fleiss (1981 ...Rates and Proportions). Dr. Weedon's
suggestion to use exact CI's is more appropriate for values of p near 0.
I can't seem to get exactly the numbers you have given, but it looks like
you have about 2 out of 115 subjects with this "rare factor".
With only 2 subjects out of 115, we can't estimate this proportion very
well at all. Is an interval estimator based on these data a good idea?
