In k fold CV how use trained model (logistic regression) to compare test data set and draw ROC

[Zinan]

Hi I am new to SAS and need suggestions on my  codes.

I wanted to compare two logistic regression models. I tried to use 10-fold cross-validation for both models. I first searched online and find below code which can generate different folds and run logistic regression on my training data sets. But I do not know how I can use the training model to predict and compare my test data set. If possible, I wanted to draw ROC and calculate the AUC.

I attached current codes here. Please help me for my next step. Thanks very much!

%let K=10;

%let rate=%sysevalf((&K-1)/&K);

*Build model with all data;

proc logistic data=cold;

model y(event="1")=temp time time*temp;

run;

*Generate the cross validation sample;

proc surveyselect data=cold out=cv seed=231258

samprate=&rate outall reps=10;

run;

/* the variable selected is an automatic variable generatic by surveyselect.*/

/*If selected is true then then new_y will get the value of y otherwise is missing */

data cv;

set cv;

if selected then new_y=y;

run;

/* get predicted values for the missing new_y in each replicate */

ods output ParameterEstimates=ParamEst;

proc logistic data=cv;

model new_y(event="1")=temp time time*temp/outroc=roc;

by replicate;

output out=out1(where=(new_y=.)) predicted=y_hat;

run;

[Reeza]

What exactly is your final model though? 

You ran a regression for each slice so you have multiple models so to speak. How are you planning to finalize that model? 

[Zinan]

Thanks very much for you reply. Links in another reply helped me find the answer (drawing ROC and calculating AUC). But since I am still new in SAS and STAT, I was wondering if you think below procedure has no problems. 

Sorry for my post seemed a little confusing. I wanted to compare two different models for their predictive power. Because the response is binary, I used logistic regression. So I used 10 fold cross validation to evaluate the two models. For each model, I used 9 folds of data to train the model and then compare the held data to draw the ROC. I was wondering AUC can be a good indicator for the evaluation of the two models. 

Thanks again!

[Ksharp · Accepted solution]

Use SCORE statement to predict test data.

https://communities.sas.com/t5/SAS-IML-Software-and-Matrix/How-to-do-k-fold-CV-with-replacements-replication/m-p/303938#U303938


Calculate AUC.

http://blogs.sas.com/content/iml/2011/07/29/computing-an-roc-curve-from-basic-principles.html

http://blogs.sas.com/content/iml/2011/06/03/a-statistical-application-of-numerical-integration-the-area-under-an-roc-curve.html

http://blogs.sas.com/content/iml/2011/07/08/the-area-under-a-density-estimate-curve-nonparametric-estimates.html

http://blogs.sas.com/content/iml/2011/05/27/obtaining-area-from-a-set-of-points-on-a-curve.html

[Zinan]

Thanks very much! Your links help me find ways to calculate the AUC. This is awesome!