Subscribe to DSC Newsletter

We are working on a model in Auto Insurance world. Need to test 2-way, 3-way interactions in data. Catch is that we only have SAS at our disposal. In all my past projects, we've typically used decision trees (CART) to solve this problem. Any ideas? Pointers towards examples, codes would be appreciated..

Tags: Interactions

Views: 147

Reply to This

Replies to This Discussion

Not 100% sure of what you are looking for, but I hope this helps,
Thanks. Havent been able to work through the contents of the paper. Will do that next week.
A standard way to to this in SAS would be to use PROC GLM and include any interaction terms in the model equation as V1*V2 for a two way interaction. Then look at the F-value to see if it is significant. If you are working with count data, like claims data, there are better tests available, and of course you can always use PROC FREQ to look at the data as contingency tables.

If you are strictly looking at decision trees, it is a little trickier. You can use TREEDISC but need to have SAS/IML installed, or you can purchase an add-on CHAID package.

Hope this helps..

-Ralph Winters
Hi Ralph,

Thats exactly our fallback plan but problem is to find out probable combinations of V1, V2.. & so on. We arent building frequency or severity models but an acquisition model.. so not many variables..
Thanks for bringing it up....

Not sure of how TREEDISC works, will read it up.
Anova with Scheffe
Try excel. Create cross tabs taking 2 variables at a time. Note the bad rate differences in each cell. Now group cells with similar (visually) bad rates. You can create 5-8 groups. Now regress by introducing dummy variables to represent these groups. For continuous variables, break them into deciles and follow the same.
I know its very crude but it is simple and works in case you want to build a model capturing interactions....let me know if this helps you at all.
Hi Amit.

When you say "...only have SAS at our disposal" I am guessing you mean you don't want to pay for other systems that would allow you to implement decision trees. You could add R to the tools at your disposal at no cost. It's easy enough to download (from CRAN) and install (on windows, UNIX and Mac platforms).

Once you have R up and running, you can use their rpart package (doesn't depend on any additional packages that have to be downloaded seperately) to build simple decision trees rather painlessly (ofcourse you can do a lot more but not quite as simply). You could use this to identify some interesting interactions, code them up and test them in your SAS model.



On Data Science Central

© 2021   TechTarget, Inc.   Powered by

Badges  |  Report an Issue  |  Privacy Policy  |  Terms of Service