Forecasting using SAS Forecast Server, SAS/ETS, and more

Model selection in proc panel

Reply
New Contributor
Posts: 2

Model selection in proc panel

Hey everyone,

I am modelling variables with a two-way fixed effects model in SAS 9.4, in proc panel.

Do you know if there is a way to use the "selection"stepwise" option in the panel procedure? And if now, do you have any tips to automatically select your variables using for example the information critera?

Thank you in advance,

F.

New Contributor
Posts: 3

Re: Model selection in proc panel

Posted in reply to flora1990

Hi Flora,

To my knowledge, PROC PANEL does not have a stepwise option. I'm not a fan of automatic selection, but there are a number of things you can do to go down that path. Given that the panel membership is reasonably highly correlated, you could pick a test subject and perform a stepwise OLS regression. Or similarly, perform a stepwise regression on the panel average. You could also produce ranked pairwise correlations to provide guidance while manually model building. On the complex side, you could write a macro for a panel stepwise selection using AIC/BIC instead of Rsq.

Hope this helps!

SAS Employee
Posts: 89

Re: Model selection in proc panel

Posted in reply to flora1990

I have been thinking about this one and here is an idea.  

1) Time and cross-sectionally demean your data.  How to do this is shown here. SAS/ETS(R) 14.1 User's Guide You will likely want to use some form of PROC TIMESERIES or some PROC MEANS calls to do this. 

2) Now that the data are "demeaned" a simple OLS estimate would give you your "Two-way FE" estimates.  From here you could then extend this regression by using PROC REG or perhaps better, PROC GLMSELECT SAS/STAT(R) 14.1 User's Guide

and use whatever variable selection method you prefer.   In essence, it would force FE into your model, with whatever final specification the algorithm chooses.

*would love to see some of your results. 

Good luck-Ken

Ask a Question
Discussion stats
  • 2 replies
  • 937 views
  • 0 likes
  • 3 in conversation