Olga Gorbachev Melloni <olga_gorbachev@hotmail.com> asks,
> I am trying to estimate a model, in which errors follow an MA(1). I
> have panel data, and I assume that for each cohort, coefficients are
> the same. I ran:
>
> arima y x if cohort==1935, ma(1) condition
>
> [...]
>
> 2. When trying to predict the errors or the predicted values, I got
> an error message saying that the sample cannot include panels.
>
> Is there a way to get the predicted values and/or white-noise
> errors?
Bob Yaffee <bob.yaffee@nyu.edu> answered Olga's first question, which I did
not repeat above.
Olga has uncovered a buglet in -arima-. No results are compromised, some
things just cannot be done directly by -arima- when estimating with the
-conditional- method, but read on, there is a work around.
This is happening because the unconditional maximum likelihood estimation
and prediction method used by -arima- is not up to panel data. Unlike the
conditional estimation method, it cannot restart itself at the beginning of
each panel. -arima- is at its heart an unconditional ML estimator, both in
implementation and as documented in the Methods and Formulas of [TS] arima.
To obtain conditional estimates -arima- actually calls through to -arch-,
which is at its heart a conditional ML estimator. -arima- then presents the
-arch- results as its own.
So long as you do not have panels, this is fine. With a single time-series,
-arima- can predict using the unconditional method of prediction, and unlike
-arch-, can gracefully handle maintaining information over any time gaps
during a prediction. (-arch- simply restarts at any gap. That is fine
asymptotically, but what -arima- does is a bit more efficient.) For that
reason, it is usually best to let -arima- predict using its unconditional
method, even when the estimates come from the -arch- conditional estimator.
We have not yet decided how to handle this in -arima-. It's not clear that
grafting the unconditional estimator into -arima- was the best approach,
though it is unlikely that Olga and others would have found it for ARIMA
modeling had it been left only in -arch-.
Regardless, until we have decided how to proceed, Olga can get her predictions
by using -arch- in place of -arima-. Just use -arch- with only ARIMA terms
and without any ARCH terms. The syntax is similar and she need not specify
-conditional- because that is the only estimator -arch- knows. The estimation
results will be identical to those obtained with -arima, conditional- and
-predict- will work on panel data after -arch-.
-- Vince
vwiggins@stata.com
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/