RESTRICT STATEMENT FOR PROC LOGISTIC

[davis]

Good morning,

I have a question, and was wondering if someone could lend a hand.

I am running a competing risk model (default risk and prepayment risk) using proc logistic via SAS EG (7.15 7.100.5.5850 64-bit).  I have one set of covariates, or right-hand side variables, for both risks.  However, I need some flexibility.  That is, I would like to restrict (i.e., omit) one covariate from this set of covariates in predicting one of the risks.  For instance, the covariate 'incentive to refinance a mortgage' is driven by mortgage rates and does not help explain default.  As a result, I would like to 'turn off' the 'incentive to refinance a mortgage' covariate in predicting mortgage default.

Currently, I am exploring the restrict statement, but I have not seen as of yet, an application of this statement applied to proc logistic.

Any assistance would be much appreciated.

[ballardw]

It may help someone that knows more about this procedure and options than I do if you post the code of what your are currently doing and then discuss things in terms of the actual variable names that appear in your code, letting use know which variable needs to be "restricted" and how.

That way someone else may be able to modify the code to get what you want.

I would think this would be two different models => model statements=> two calls to proc logistic.

[SteveDenham]

Part 1: I can't find anything in the documentation about a RESTRICT statement for PROC LOGISTIC, so anything like that has to be handled through options, like defining an offset function..

Part 2: Part 1 probably will run into problems defining the coefficient as zero, as the offset is fixed for each observation (coefficient=1).

About the only way I can think of doing this is to use something like PROC NLIN or PROC NLMIXED where you can program around this.  It would take some time to kludge up something, as the fitting of two models where one drops an effect is pretty easy, but welding the two results together within a single matrix framework to enable testing seems difficult for me.

SteveDenham

[StatDave]

Why not simply remove that variable from the model for default? Assuming that's what you mean by "turn off". If you can to fix its coefficient to a particular value, c, then create a variable that is c*incentive and specify it in the OFFSET= option. See this note for general information on restricting some or all model predictors in many procedures.