Re: [Gretl-users] Sequentially dropping insignifcant regressors
Allin Cottrell schrieb:
>
> -----
> Hi, I have just discovered gretl. It is quite convenient that
> insignificant regressors can be sequentially dropped from a
> regression. However, sometimes want does not want to drop certain
> regressors right away, even if they are insignificant (e.g. variables
> in levels in an unrestricted error correction model). Is it possible
> to exclude some variables from the routine? How?
You can do this quite easily via a script. I'm not sure it's
worth implementing as a built-in command.
sorry Allin, I don't see how, at least not in combination with the
built-in auto omit feature. (It's clear that a handmade script can be
written for a specific situation, but then it may be no quicker than
doing the sequential drop by hand.)
As a workaround, one could do:
1. omit=auto
2. in case any of the interesting variables have been dropped, re-insert
them at the end (ideally just for documentation purposes, but it could
happen that they suddenly appear as nominally significant)
Another workaround could be:
1. estimate the full model x0 = b1*x1 + b2*x2 + b3*x3 ....
2. construct a new left-hand-side variable by moving the variables to be
retained to the left-hand side of the equation, using the provisionally
estimated coefficients from step 1, say x0 - b1hat*x1 = b22*x2 + b33*x3 ...
3. do omit=auto for this transformed model to filter out the remaining
important regressors, say it turns out that only the coefficient of x3
is significant
4. then estimate as the final model x0 = b111*x1 + b333*x3
Both workarounds are not perfect of course and can cause problems.
OTOH, users should not expect this feature to be a replacement for a
full-blown general-to-specific search a la "PcGets" or "Autometrics".
That would involve diagnostic tests at each step for example.
BTW, I think in Jmulti you can do such a subset search with some terms
marked as never to be deleted.
good luck everyone,
sven