topic Re: Calculation about score function in Proc Logistic in Statistical Procedures

Calculation about score function in Proc Logistic

fannavivian — Tue, 21 Feb 2017 20:52:53 GMT

Hi,

Can anyone tell me how SAS calculates probalibility using proc logistic score statement? Why my manual probability calculation is so different from SAS's? Does SAS take all variables into calculation or only significant ones?

P=exp(F)/(1+exp(F)) = 1 / (1+exp(-F))

Analysis of Maximum Likelihood Estimates Parameter DF Estimate Standard
Error Wald
Chi-Square Pr > ChiSq

Intercept		1	-0.9724	0.2331	17.4036	<.0001
classA	1	1	0.0150	0.0827	0.0328	0.8562
classA	2	1	-0.0694	0.0380	3.3468	0.0673
classA	3	1	-0.1681	0.0467	12.9276	0.0003
classA	4	1	0.1498	0.0706	4.5050	0.0338
classB	1	1	-0.1760	0.0547	10.3536	0.0013
classC	1	1	1.4962	0.0280	2863.0571	<.0001
classD	1	1	-0.5437	0.0445	149.1918	<.0001
classE	1	1	0.1358	0.1794	0.5727	0.4492
classE	2	1	0.1191	0.3248	0.1345	0.7138
classE	3	1	-0.3680	0.4482	0.6740	0.4117
Numeric1		1	-0.2314	0.0197	138.3852	<.0001
Numeric2		1	0.000259	0.000040	42.9420	<.0001
Numeric3		1	2.1190	0.1641	166.7216	<.0001

Thank you!!!

Fan

Re: Calculation about score function in Proc Logistic

Reeza — Tue, 21 Feb 2017 21:01:12 GMT

It uses the variables listed in the ParameterEstimates tables.

If you're doing a selection method, the last table is the significant one.

If you want code that replicates the scoring code, look at CODE statement within PROC LOGISTIC.

See the various methods for scoring data here:

http://blogs.sas.com/content/iml/2014/02/19/scoring-a-regression-model-in-sas.html

Here's a fully worked example to demonstrate the issue.

https://communities.sas.com/t5/SAS-Statistical-Procedures/How-to-determine-logistic-regression-formula-from-estimates/m-p/120780/highlight/true#M6321

Re: Calculation about score function in Proc Logistic

fannavivian — Wed, 22 Feb 2017 16:35:05 GMT

Appreciate your answer. Please correct me if I am wrong. As for SAS's calculation, it takes all the variables in the parameter estimates report, just the all the predictor variables. And as in your answers, there's still difference. So I really would like to know what makes the difference.

Thanks!

Re: Calculation about score function in Proc Logistic

Rick_SAS — Wed, 22 Feb 2017 17:00:27 GMT

Without seeing your code or data, it is difficult to guess your problem.

Do you have missing values? If so, you need to handle them correctly. Are you using the full precision of the parameter estimates or just the 4 decimal place shown in the table?

The following is a brief example that uses the DATA step to score. In practice, you would use the CODE statement, the SCORE statement, or PROC PLM, as explained in the article that Reeza links to.

proc logistic data=sashelp.class;
model sex = height weight age;
output out=LogiPred(keep=P) pred=P;
run;

data DataPred;
set sashelp.class;
F = -1.19262669802445  + 
    -0.15172772458382 * Height +  
    -0.13262689324423 * Weight +
     1.76465589368491 * Age;
P = 1 / (1 + exp(-F));
keep P;
run;

proc compare base=LogiPred compare=DataPred method=absolute criterion=1e-12;
run;

Re: Calculation about score function in Proc Logistic

Reeza — Wed, 22 Feb 2017 18:18:31 GMT

99% of the time the difference will be due to an error in the manual calculation.