Multinomial logistic regression with different driver variables

[mphrna1989]

I am attempting to build a multinomial logistic model that predicts various states of a loan's performance (ie., Current, Past Due, Defaulted, Paid off). Although proc logisitc would seem like a good candidate, the issue is that prepayment has very different drivers than default/delinquency. Even though I could use a variable selection procedure, it still selects undesired variables. Ideally, I would prefer to have different sets of variables for prepayment, while preserving the joint modeling. Is there a way to do this? 

Thanks for any assistance you may provide.

[PaigeMiller]

the issue is that prepayment has very different drivers than default/delinquency

Regardless of what you are modelling, or what modelling method you use, this sounds to me like a perfect and justifiable reason to use different models.

[mphrna1989]

Thanks for the feedback. The advantage of modeling the events jointly is that I could account for the correlation between paying off and defaulting. Of particular concern is that one who is likely to payoff is extremely unlikely to default. To your point, I do currently model them separately, and I do certainly see the validity in that argument.

[PaigeMiller]

If you are looking to understand and interpret the model that has been fit, I can't see one model giving good interpretability, because as you said, the drivers are different.

If you just want a model that predicts well, but has low interpretability, then I can see one model working.

[Reeza]

I think you need to consider context here, which we don't have. 

Is pre-payment the one where a person has to pre-load a credit card with X amount and then borrows against it, or is it people who pre-pay their loans or pay higher amounts? 

Depending on what you're trying to do, perhaps you can use a stratification approach (STRATA) for the loan type? 

If it's the first case, I would say they don't fit into regular loans because the bank has already identified them as high risk. 

If it's the second, I would lean to including that as a factor/categorical variable in the model.

[mphrna1989]

Prepayment is when a borrower pays off their loan. The primary driver here would be their cost savings (ie., refinancing and getting a lower rate). So, essentially, a loan can terminate by prepaying or defaulting. Thanks.