Prev: 2 PART MODEL
Next: A UK news outlet picks up on the SAS vs. WPS suit
From: raj on 22 Nov 2009 23:57
Hi,

I am running a logistic regression with a binary response variable
and many binary, continuous predictor variables and a treatment effect
in my model. I have 2 treatment groups and i wish to get the "AVERAGE"
PREDICTED PROBABILITY FOR EACH TREATMENT GROUP. How can i get it using
a sas code.? Any help on this would be highly appreciated?

Rj
From: Dirk Nachbar on 23 Nov 2009 08:23
On Nov 23, 4:57 am, raj <rkamalaka...(a)gmail.com> wrote:
> Hi,
>
>   I am running a logistic regression with a binary response variable
> and many binary, continuous predictor variables and a treatment effect
> in my model. I have 2 treatment groups and i wish to get the "AVERAGE"
> PREDICTED PROBABILITY FOR EACH TREATMENT GROUP. How can i get it using
> a sas code.? Any help on this would be highly appreciated?
>
> Rj

you do

proc logistic;
model y=x1 x2 treatment;
output out=p p=prob;
run;

proc means;
class treatment;
var prob;
run;

Dirk Nachbar
From: Dale McLerran on 24 Nov 2009 00:45
--- On Mon, 11/23/09, Dirk Nachbar <dirknbr(a)GOOGLEMAIL.COM> wrote:

> From: Dirk Nachbar <dirknbr(a)GOOGLEMAIL.COM>
> Subject: Re: Logistic regression
> To: SAS-L(a)LISTSERV.UGA.EDU
> Date: Monday, November 23, 2009, 5:23 AM
> On Nov 23, 4:57 am, raj <rkamalaka...(a)gmail.com>
> wrote:
> > Hi,
> >
> > I am running a logistic regression
> with a binary response variable
> > and many binary, continuous predictor variables and a
> treatment effect
> > in my model. I have 2 treatment groups and i wish to
> get the "AVERAGE"
> > PREDICTED PROBABILITY FOR EACH TREATMENT GROUP. How
> can i get it using
> > a sas code.? Any help on this would be highly
> appreciated?
> >
> > Rj
>
> you do
>
> proc logistic;
> model y=x1 x2 treatment;
> output out=p p=prob;
> run;
>
> proc means;
> class treatment;
> var prob;
> run;
>
> Dirk Nachbar
>

This approach would produce estimates of mean values
conditional on the observed predictor variable distribution.
That is probably not what is desired. Suppose that the
different treatment conditions experienced different
distributions of the adjustment variables. Then the
predicted probabilities for each treatment condition
are based on characteristics of the adjustment variables
in addition to the treatment condition itself.

Instead, I would suggest using the GLIMMIX procedure
and generating least squares means back transformed to
the original (inverse link function) scale. The code
would be something like

proc glimmix data=mydata;
class treatment x1 x2;
model y(ref=first) = treatment x1 x2 x3 x4 / dist=bin;
lsmeans treatment / om ilink;
run;


Note that the model statement specifies that the first
level of the response will be the employed as the
reference level meaning that for the binary response,
the other level would be the event of interest. The
LSMEANS statement indicates that we want mean values
for the treatment variable on the inverse link scale
and computed at the observed marginal (OM) distribution
of the other predictor variables.

Dale

---------------------------------------
Dale McLerran
Fred Hutchinson Cancer Research Center
mailto: dmclerra(a)NO_SPAMfhcrc.org
Ph: (206) 667-2926
Fax: (206) 667-5977
---------------------------------------
 | 
Pages: 1
Prev: 2 PART MODEL
Next: A UK news outlet picks up on the SAS vs. WPS suit