Please define 'mean probabilities'.
To compute the C-index or Dxy you need anything that is monotonically
related to the prediction of interest, including the linear combination
of covariates ignoring all intercepts. In other words you don't need
to go to the trouble of computing probabilities unless you are binning,
as the binning is usually done on a controllable 0-1 scale. When I bin
I just choose the middle intercept, I seem to recall. Also try running
SAS with a very tiny BINWIDTH and see if you get 1 - .968 as the answer
for C. [I wrote the original algorithm SAS uses for this in the old SAS
PROC LOGIST. Binning was just for speed.]
You might also re-run SAS after negating the response variable.
Frank
blackscorpio wrote
Dear Dr Harrell,
Thank you very much for your answer. Actually I also tried to found the C
index by hand on these data using the mean probabilities and I found
0.968, as you just showed.
I understand now why I had a slight difference with the outpout of lrm. I
am thus convinced that this result is correct.
I read on the SAS help that the procedure logistic also proceed to some
binning (BINWIDTH option) :
http://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#statug_logistic_sect010.htm
But I cannot explain why the difference between the two softwares is that
huge, especially since the class probabilities are the same.
Do you think it could be due to the fact that mean probabilities are
computed differently ?
Thank for your help and best regards,
OC
Date: Thu, 24 Jan 2013 05:28:13 -0800
From:
f.harrell@
To:
r-help@
Subject: Re: [R] Difference between R and SAS in Corcordance index in
ordinal logistic regression
lrm does some binning to make the calculations faster. The exact
calculation
is obtained by running
f <- lrm(...)
rcorr.cens(predict(f), DA), which results in:
C Index Dxy S.D. n
missing
0.96814404 0.93628809 0.03808336 32.00000000
0.00000000
uncensored Relevant Pairs Concordant Uncertain
32.00000000 722.00000000 699.00000000 0.00000000
I.e., C=68 instead of .963. But this is even farther away than the
value
from SAS you reported.
If you don't believe the rcorr.cens result, create a tiny example and do
the
calculations by hand.
Frank
blackscorpio81 wrote
> Dear R users,
>
> Please allow to me ask for your help.
> I am currently using Frank Harrell Jr package "rms" to model ordinal
> logistic regression with proportional odds. In order to assess model
> predictive ability, C concordance index is displayed and equals to
0.963.
>
> This is the code I used with the data attached
> data.csv <http://r.789695.n4.nabble.com/file/n4656409/data.csv>
> :
>
>>require(rms)
>>a<-read.csv2("/data.csv",row.names =,na.strings = c(""," "),dec=".")
>>lrm(DA~SJ+TJ,data=
>
> Logistic Regression Model
>
> lrm(formula =A~SJ+TJ, data = a)
>
> Frequencies of Responses
>
> 1 2 3 4
> 6 13 9 4
>
> Model Likelihood
> Discrimination Rank Discrim.
> Ratio Test
> Indexes Indexes
> Obs 32 LR chi2 53.14
R2
> 0.875 C 0.963
> max |deriv| 6e-06 d.f. 2 g
> 8.690 Dxy 0.925
> Pr(> chi2) <0.0001
gr
> 5942.469 gamma 0.960
>
> gp 0.486 tau-a 0.673
>
> Brier 0.022
>
> Coef S.E. Wald Z
Pr(>|Z|)
> y>= -0.6161 0.6715 -0.92 0.3589
> y>= -6.5949 2.3750 -2.78 0.0055
> y>= -16.2358 5.3737 -3.02 0.0025
> SJ 1.4341 0.5180 2.77 0.0056
> TJ 0.5312 0.2483 2.14 0.0324
>
> I wanted to compare the results with SAS. I found the same slopes and
> intercept with opposite signs, which is normal since R models the
> probabilities P(Y>=X) whereas SAS models the probabilities P(Y<=k|X)
> (see pdf attached, page 2 , table "Association des probabilités
prédites
> et des réponses observées").
> SAS_Report_-_Logistic_Regression.pdf
>
<http://r.789695.n4.nabble.com/file/n4656409/SAS_Report_-_Logistic_Regression.pdf>
>
> I chose the order for levels.
>
> I controlled that the corresponding probabilities P(Y=X) are the
same
> with both softwares. But I can't understand why in SAS the C index
drops
> from 0.963 down to 0.332.
>
> I read a lot of things about this and it seems to me that both
softwares
> use slightly different technique to compute the C index ; it is
> nevertheless surprising to me to observe such a shift in the results.
>
> Does anyone have a clue on this ?
> Thank you very much for you help
> Blackscorpio
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.