Scoring new data - proc LOGISTIC with OFFSET option

[Lahrie]

Hi,

I used PROC LOGISTIC with OFFSET option, as I'm dealing with rare target data. Is it possible to score a new dataset without using inmodel option? How to do it?

The new dataset will be scored using another program, so I can't use any SAS function to score it.

Tks

[PaigeMiller]

The new dataset will be scored using another program, so I can't use any SAS function to score it.

Explain this. Why are you asking a SAS question and then say you can't use SAS to do the scoring.

[StatDave]

See this note. As shown there, predicted probabilities can be produced by removing the offset from the x*beta computation and then applying the inverse link function: 1/1+exp(-xbeta) .  The inverse link function is implemented as the LOGISTIC function in the DATA step which is used in the note. 

[Reeza]

Use the CODE statement within PROC LOGISTIC to get the code, check the documentation for details. You'll still have to do some 'translation' but it's much easier. If you're using SAS Enterprise Miner it can give you code in several languages.

[Lahrie]

Hi. I used the CODE option and I saw that is necessary to inform the offset value. My new data have the real proportion of goods and bads. How I can get the code without using the offset value?

Tks

[StatDave]

If you were going to do the scoring in SAS, I would suggest that you just write a DATA step to compute x*beta omitting the offset and then apply the inverse link as I mentioned. You can see some example code in section 4 of this note. But since you say you are not going to use SAS to do the scoring, then you will have to do the equivalent of that in whatever you intend to use. Still, you might find it helpful to first do it in SAS and satisfy yourself that you get the correct scores before translating that into whatever you use to score.

[Lahrie]

Mar 17, 2020 1:16 PM (41 views)  |  Posted in reply to message from Lahrie 03-16-2020

If you were going to do the scoring in SAS, I would suggest that you just write a DATA step to compute x*beta omitting the offset and then apply the inverse link as I mentioned. You can see some example code in section 4 of this note. But since you say you are not going to use SAS to do the scoring, then you will have to do the equivalent of that in whatever you intend to use. Still, you might find it helpful to first do it in SAS and satisfy yourself that you get the correct scores before translating that into whatever you use to score.

Hi @StatDave ,

I've tested your suggestion but I couldn't get the correct prob. Let me describe my steps:

proc logistic data=datax;
model conc(event="0")= SU FIN gp5
/selection=stepwise offset=off;
run;

This logistic give me the following estimates:

Parameter Estimate
Intercept -7.212
SU          0.7096
FIN         1.5471
GP5        0.9827
off           1.000

So, I tried to score a new dataset:

data sample;
input SU FIN GP5;
datalines;
0 0 0
1 1 0
1 0 0
1 1 1
0 0 1
;

data z;
set sample;

xbeta= - 7.2120 +
    SU * 0.7096 +
   FIN * 1.5471 +
   GP5 * 0.9827 ; /* I've ommited the offset estimate */

p = 1 / (1 + exp(-xbeta));
run;

results below:
SU FIN GP5 xbeta p
0 0 0 -7.212 0.0007371365
1 1 0 -4.9553 0.0069966679
1 0 0 -6.5024 0.0014975891
1 1 1 -3.9726 0.0184766143
0 0 1 -6.2293 0.0019669545

the P is too much small and very different from those score in logistic step. Could you help me?

Tks

[Lahrie]

If you were going to do the scoring in SAS, I would suggest that you just write a DATA step to compute x*beta omitting the offset and then apply the inverse link as I mentioned. You can see some example code in section 4 of this note. But since you say you are not going to use SAS to do the scoring, then you will have to do the equivalent of that in whatever you intend to use. Still, you might find it helpful to first do it in SAS and satisfy yourself that you get the correct scores before translating that into whatever you use to score.

[StatDave]

Using the example in the first note I referred to (22601: Adjusting for oversampling the event level in a binary logistic model), these statements compute (in POFF2) essentially the same values as those computed using the logistic function after removing the offset. They differ slightly because the POFF2 computation does not use the full precision of the parameter estimates.

      proc logistic data=sub;
        model y(event="1")=x / offset=off;
        output out=out xbeta=xboff;
        title2 "Offset-adjusted Model";
        run;
      data out;
        set out;
        poff=logistic(xboff-off);
        poff2=1/(1+exp(-( -3.4596 + 2.3149*x )));
        run;