Rolling Regression over Series

[yelena]

Dear Tom,

I have eight variables in each case and three options (lags) for each variable. So together I may have 3^7*2 = 4374 cases (2 is because DEPENDENT variable can only have 2 choices, DEPENDENT{1}, DEPENDENT{2}). Out of 4374 cases, I have to select the model with high R^2 and significant variables. Now, if all three lags of the same variable are insignificant, I have to keep the lag with the lowest p value, even if p>0.05. I cannot drop any varaibles from the model. I have to keep them all. But lags may change.

I looked at your rollreg.src file but it seems that it rolls the regression over time period. Is there any tool that allows to select the most significant lags without dropping all varaibles. STWISE function, does it have any additional option like that?

Thank you.

[TomDoan]

To be perfectly honest, I've never heard of such a thing. I'm a bit confused about your choices for lags. You do {1}, {1 2}, {1 2 3} as long as all are significant, but could have {2} or {3} if none of the others yield all significant coefficients? If that's correct, then the number of possible regressions is quite a bit higher, by a factor of about 40. This, BTW, is a completely different analysis than a rolling regression, which is (by contrast) quite simple. I don't think there's an alternative to just crunching out every regression, since changing lags on own variable will change the significance of lags on others.

[yelena]

Okay, I agree that adding the lags will change the coefficients of others. But is there an automatic way to filter a number of possible regressions, in which we will have all variables used with a different combination of lags?

[TomDoan]

I don't think you've come up with a workable criterion---it's possible that there is no set of lags which meets your standards for all variables. Is there a reason that you can't force lag one on all variables and then check only the lags 2 and 3?