I must be the one who is advocating against the bootstrap on this list
most aggressively :)).
My personal advice would be to sample a panel as a whole, thus
preserving its unique dependence structure. I don't think I've seen
this properly justified anywhere though. You can do this with
-bsample, cluster(id) idcluster(newid)-, and then run your GMM command
as -xtabond ... , i(newid)-. One thing you would certainly lose with
this approach is the control over your sample size. The trick then
might be this:
bysort id (time): gen byte panelsize = _N
bsample, cluster(id) idcluster(newid) strata(panelsize)
That way, Stata will resample the same number of clusters (panels)
from each stratum (defined as panels of the same size), and thus your
sample size and the pattern of unbalancedness will be preserved.
If you had a complex survey sample design on top of the panel
structure, you would have to forget about control over the sample
size, as you would need to sample within the strata, and take a bit
fewer clusters than there were in the original sample; see Rao, J. N.
K. & Wu, C. F. J. (1988), 'Resampling inference with complex survey
data', J. Amer. Stat. Assoc. 83, 231– 241.
Stas
On Thu, 4 Nov 2004 18:58:48 -0000, Nick Cox <[email protected]> wrote:
> You should search the Statalist archives. There have many
> posts in the last few months on bootstrapping
> time series. The consensus seems to be that
> naive bootstrapping with time series is a very
> bad idea, for precisely the reason you mention,
> dependence structure. Special methods have been proposed
> to bootstrap time series, but none appears to
> have been implemented in Stata. On the other hand,
> bootstrapping in terms of selecting or not selecting
> panels may not be so crazy, but others may well
> comment in more detail and with more authority.
>
> Nick
> [email protected]
>
> Carmine Ornaghi
> >
> > I have an unbalanced panel data of firms (with
> > observations between 3 and 10 periods). I use this
> > data to estimate a production function using dynamic
> > GMM (Arellano and Bond).
> >
> > Now I want to do some sort of 'bootstrapping. In
> > particular, I have the following doubts:
> > 1) I think I need to preserve the same time structure.
> > What I mean is that I want the random sample to have
> > the same number of firms with only 3 observations, 4
> > observations, ... 10 observations that the original
> > dataset has. In this way they are UNBALANCED in the
> > same way
> > 2) I have done some experiment with the STATA command
> > bsample but I have a problem. The Sargan Test of
> > overidentified restriction is easily passed with the
> > original dataset but never with all the random samples
> > created. Do you know why and how I can give a higher
> > probability of entering into the sample to those
> > observation that minimize the Sargan Test.
> >
> > Thanks a lot,
> >
> > Carmine
> >
> >
> >
> >
> > ___________________________________
> > Nuovo Yahoo! Messenger: E' molto più divertente: Audibles,
> > Avatar, Webcam, Giochi, Rubrica... Scaricalo ora!
> > http://it.messenger.yahoo.it
> > *
> > * 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/
> >
>
> *
> * 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/
>
--
Stas Kolenikov
http://stas.kolenikov.name
*
* 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/