A prospect is evaluating JMP Pro and they ran into a JMP Alert warning using the Fit Model -> Nominal Logistic regression personality that at first flummuxed us both but upon further review we figured it out. Lesson learned. In part, if nothing else, this little narrative points out the value of using platforms such as Explore Missing Values and Missing Data Pattern, among others, BEFORE modeling to evaluate the appropriateness of a data set for the intended modeling and practical purposes.
On first run through the modeling work...the customer has a 2 level, nominal response. In essence a pass/fail response (banking credit worthiness type problem). About 50 predictor variables, a mix of continuous and nominal modeling type x's. They decided to throw all the x's into the Construct Model Effects window in one fell swoop as main effects only...and hit run. A JMP Alert message populated alarming that the response needs more than one level. They then tried a few alternative models such as fitting only the continuous predictors...everything 'worked' as intended. Then they came calling me asking "Why do I get this alert...and JMP stops working for some models but not others?"
I pondered this a bit and what we discovered was the union set of all the missing value rows among all the predictors created a condition where in essence one complete level of the response was excluded from the fit because of the missing values.
Made perfect sense once we spent some time on this...thought I'd share if you ever run into something like this. And if nothing else this spawns additional conversation on using JMP Pro for data viz BEFORE analysis...and then various ways to deal with missingness...informative missing, imputation, etc.
Ghost data anyone?
Yes! Please do write this up as a blog post! I can do a quick call with you to discuss and explain the process.
Thanks, Peter. You're right these are good lessons and would make a great blog post. @arati_mejdal can point you in the right direction if you need help writing it.
