Proc Reg Forcing Variable

[Giraffe123]

Hi Everyone,

I'm trying to fit a regression model in sas to a default series using macroeconomic variables. I have 6 potential macro variables. 3 of these should be forced to be in the model (gdp, cpi and prime). 

I have 24 lags on each of  the variables (i.e. gdp lag1, gdp lag2 .... gdp lag 24). I would like to force the 3 core variables (gdp, cpi, prime) or any of their variables (lag1 - lag 24) into the model while testing the remaining 3 variables (hddi, hpi and vsr) to get the best fit. I'm allowing for 6 variables in the model.

Thus, I'm looking for a potential model that looks like this (for example)

def_rate = gdp_lag18 + prime_lag2 + cpi_lag6 + hddi_lag12

or

def_rate = gdp_lag8 + prime_lag15 + cpi_lag8 + hddi_lag2 + hpi_lag1

etc.

Point is I would like to force the proc regression model to include the 3 core macrovariables and any of their variants. Does anyone know an easy way to do this?

I'm aware of the include= function, but I don't think you can specify either/or in the option (for instance it can be GDP_lag 2 or GDP_Lag15 ,but must include GDP, prime and cpi)

I know one can also output every model possible by not specifying in the best =  function, and then filter on your output, just don't think this is the optimal method because you will be fitting thousands of models.

Thanks!

[PaigeMiller · Accepted solution]

You might want to try the GROUPNAMES= option in MODEL statement of PROC REG. That may or may not do what you are asking.

A better solution, in my opinion, is to fit a Partial Least Squares model (PROC PLS) and then all of your variables are included in the model, in such a way that the effect of multi-collinearity is reduced. See here for an example of a PLS model which had 1000 highly correlated predictor variables and used them all and produced a usable model. One benefit is that all this effort to select variables to put into the model is eliminated or reduced.