PROC ADAPTIVEREG, UNBALANCED PANEL, BEHAVIOR OF LOAN BORROWER

[davis]

I am trying to determine whether Proc Adaptivereg is a reasonable analytic tool to estimate the probability of behavior of a loan borrower (e.g., prepay, default, remaining current) in a future period. The estimation sample is reasonably large (i.e., sample size > 3 million) and is an unbalanced panel data set where the number of rows (i.e., monthly loan performance data) exceed the number of columns (i.e., explanatory factors). In addition, practitioner literature suggests a multinomial logistic approach to model this non-linear relationship.

Lastly, I am using SAS EG - 7.1, SAS/Base - 9.04, and SAS/STAT - 14.2.

Suggestions would be much appreciated.

[Rick_SAS]

I think @StatsMan misspoke. You can use the DIST= option on the MODEL statement in PROC ADAPTIVEREG to fit discrete responses such as binomial, negative binomial, and Poisson. For an example that analyzes a binary response, see "Nonparametric regression for binary response data in SAS." 

ADAPTIVEREG is a nonparametric procedure. For fitting parametric models, you can use the procedures (LOGISTIC or HPLOGISTIC) that StatsMan mentions, or use PROC HPGENSELECT. For a multinomial response, PROC HPGENSELECT support logit and generalized logit links, depending on whether the response if ordinal or nominal. Y

[davis]

Thanks.



One more question, if I may.



I was under the impression the outcome variable in proc adaptivereg could also be categorical.



Am I misinformed?



See the screen shot below (Source: Kuhfeld and Cai. Introducing the New ADAPTIVEREG Procedure for Adaptive Regression. SAS Global Forum 2013, Paper 457-2013, p. 11)



[cid:image001.png@01D48AF0.37869CE0]

[Rick_SAS]

A categorical variable can be ordinal or nominal. So I just used different words.

[davis]

Thanks!