topic Re: Derived Variable in Statistical Procedures

Derived Variable

Ujjawal — Tue, 23 Aug 2016 07:32:24 GMT

I am building a retail customer attrition model for saving accounts of a bank. I am targeting both soft and hard attrition (Dependent Variable).
By 'soft attrition', i mean if saving balance drops by 80% or more. By 'hard attrition', i mean customer closes his account.
My question - can i use % balance change as a predictor/ independent variable in the model. Since i am already taking customers having 80% or more balance dropped as an event in the dependent variable, would it be right to use the % change as an independent variable in the model?

Re: Derived Variable

Ksharp — Tue, 23 Aug 2016 07:51:21 GMT

What kind of model are you using ?
If it is logistic model , I think you can use % balance change as Y .
For example:

data have;
 set have;
retain n 1;
 per_balance_change=.........;
run;

proc logistic ...........
 model per_balance_change/n = ...........;

You can take HARD as per_balance_change=1  .

Re: Derived Variable

Reeza — Tue, 23 Aug 2016 10:04:28 GMT

As long as the information predates your knowledge of their attrition. So perhaps % change in prior month?

Using it over same period as your definition of dependent would be incorrect as then this is Informatiom is after the fact.

Re: Derived Variable

Rick_SAS — Tue, 23 Aug 2016 10:14:43 GMT

80% drop from where? Maximum all time balance? Previous three months? As @Reeza says, you might want to introduce a lagged variable (3 or 6 months?) to help you formulate this model.

Re: Derived Variable

Ujjawal — Tue, 23 Aug 2016 12:58:11 GMT

Thanks everyone for your response. @Rick_SAS and @Reeza - Yes, it's 80% drop from maximum all time balance. I have considered it as a target variable. Can i use % balance change from maximum all time balance as an independent variable? If i am not wrong, you are suggesting i can use % change in 3 months or 6 months as independent variables, not % balance change from maximum all time balance.

Re: Derived Variable

Rick_SAS — Tue, 23 Aug 2016 13:05:36 GMT

SAS will enable you to build either model. I don't have any intuition about which will perform best, so I will leave that modeling decision to you. You can test multiple models on historial data and choose the one that performs the best.

Re: Derived Variable

Reeza — Tue, 23 Aug 2016 13:15:57 GMT

If it's too relate to your definition of attrition it won't make logical sense. It'll make your model seem more accurate than it actually is..