Variable Screening for Multinomial Logistic Regression on Very Large Data Sets as Applied to Direct Response Modeling

Beginning with version 8.2 SAS supports multinomial logistic regression as part of PROC LOGISTIC. As a direct response modeling provider, we have found that multinomial logistic regression models often provide us with our best solutions. However, using multinomial logistic regression presents some challenges. We are often faced with very large sample sizes and a large number of candidate predictor variables. The sheer sizes of our data sets make variable selection and scoring significant issues. The functionality that SAS has built into PROC LOGISTIC to handle variable selection and scoring is not feasible for the size of data sets that we use. I will present how we have addressed the issue of variable screening in our modeling environment. I will provide a comparison of gains charts that show that the multinomial logistic model is preferable to other solutions. I will discuss and share bits of code on how we utilize univariate screening, resampling methods, and clustering for variable reduction. Discussing the issues regarding creating efficient production scoring modules utilizing multinomial logistic regression is beyond the scope this presentation. Upon request I will supply code for a simple macro that creates a text file version of a coefficient file that can be incorporated into an efficient production scoring environment.