Prev: help with counting occurences of a string
Next: Hostname
From: nuria on 9 Feb 2010 10:24
Hi,

I have a simple model were y is a binomial variable (1= sick, 0=not
sick) and x is a continous explanatory variable (for instance,
weight):

proc logistic desc;
model y = x/ctable pprob = (0 to 1 by .10) outroc=roc rsq clodds=wald;
run;


Question 1: How can I know the x value associated to each prob level,
that is, associated to each combination of sensitivity and
specificity? It is very important to know to set clinical cut-off
points.

Question 2: How can I get the 95%CI for the AUC (called "c" in the
output)?

Thanks for your help!
From: "Dennis G. Fisher, Ph.D." on 9 Feb 2010 12:32
Nuria,
You ask a good question. We have had a similar problem and had to do this
in a brute force method. What we have found helpful is a macro that was a
poster at the SUGI in San Antonio that was done by people from Eli Lilly.
Please post any answers that you get to the list. Thank you.
Dennis Fisher

Dennis G. Fisher, Ph.D.
Professor and Director
Center for Behavioral Research and Services
California State University, Long Beach
1090 Atlantic Avenue
Long Beach, CA 90813
tel: 562-495-2330 x121
fax: 562-983-1421


-----Original Message-----
From: SAS(r) Discussion [mailto:SAS-L(a)LISTSERV.UGA.EDU] On Behalf Of nuria
Sent: Tuesday, February 09, 2010 7:25 AM
To: SAS-L(a)LISTSERV.UGA.EDU
Subject: ROC curves

Hi,

I have a simple model were y is a binomial variable (1= sick, 0=not
sick) and x is a continous explanatory variable (for instance,
weight):

proc logistic desc;
model y = x/ctable pprob = (0 to 1 by .10) outroc=roc rsq clodds=wald; run;


Question 1: How can I know the x value associated to each prob level, that
is, associated to each combination of sensitivity and specificity? It is
very important to know to set clinical cut-off points.

Question 2: How can I get the 95%CI for the AUC (called "c" in the output)?

Thanks for your help!
From: Robin R High on 9 Feb 2010 13:01
Nuria,


1) Since you can save the probabilities in the outroc = data file and also
the model coefficients with ODS, you can compute the X values backwards:

ods output parameterestimates=prms;

proc logistic data=example ;
model y = x / outroc=ROCData;
run;

proc print data=prms NOobs; run;

proc transpose data=prms output=coef(drop=_name_) prefix=_;
var estimate; id variable;
run;

proc print data=coef NOObs; run;

* find the x values associated with each predicted value;

data rocdata; set rocdata; if _n_ = 1 then set coef; drop _intercept _x;
x_values = (LOG(_prob_/(1-_prob_))- _intercept)/(_x);

proc print data=rocdata(obs=15) NOobs ; run;



for 2) the association table is found:

ods select association;
proc logistic data=example;
model y=x;
run;

You can also compute auc (the c value) with PROC FREQ and then compute the
asymptotic standard error and conf intervals in a DATA step:

ODS OUTPUT measures=msrs(KEEP=Statistic Value ASE);

PROC FREQ DATA=example;
TABLE y * x / measures;
run;

DATA msrs; SET msrs;
IF LOWCASE(SUBSTR(statistic,11,3)) eq 'c|r' THEN
DO;
auc = (value/2) + .5;
se_auc = ase/2;
auc_upper = auc + 1.96 * se_auc;
auc_lower = auc - 1.96 * se_auc;
output;
end;

PROC PRINT DATA= msrs; run;


This latter computation is described in "Analyzing Rec Oper Char Curves
with SAS" by Mithat Gonen, Chapter 3

There are also bootstrapping routines to compute this interval described
in chapter 3.9

Robin High
UNMC







From:
nuria <nchapinal(a)YAHOO.COM>
To:
SAS-L(a)LISTSERV.UGA.EDU
Date:
02/09/2010 11:12 AM
Subject:
ROC curves
Sent by:
"SAS(r) Discussion" <SAS-L(a)LISTSERV.UGA.EDU>



Hi,

I have a simple model were y is a binomial variable (1= sick, 0=not
sick) and x is a continous explanatory variable (for instance,
weight):

proc logistic desc;
model y = x/ctable pprob = (0 to 1 by .10) outroc=roc rsq clodds=wald;
run;


Question 1: How can I know the x value associated to each prob level,
that is, associated to each combination of sensitivity and
specificity? It is very important to know to set clinical cut-off
points.

Question 2: How can I get the 95%CI for the AUC (called "c" in the
output)?

Thanks for your help!
 | 
Pages: 1
Prev: help with counting occurences of a string
Next: Hostname