Re: Re: st: Bootstrap with xtregar fails

[Austin Nichols <austinnichols@gmail.com>]

Nick--
Also "both T and N are
large" so there is some confusion on the part of the OP.
T is the # of time periods, and N is the number of panels AKA clusters
for this case.

On Tue, Jan 15, 2013 at 10:48 AM, Nick Cox <njcoxstata@gmail.com> wrote:
> But Vasja said that he has just one panel...
>
> Nick
>
> On Tue, Jan 15, 2013 at 3:32 PM, Austin Nichols <austinnichols@gmail.com> wrote:
>> wasis dat <vasja.sivec@gmail.com>:
>> -bootstrap- does acknowledge the error dependence if you resample
>> clusters instead of obs, but you need to tell -xtregar- what to do
>> about resampled clusters. See e.g.
>> http://www.stata.com/statalist/archive/2006-11/msg00025.html or
>> http://www.stata.com/statalist/archive/2007-08/msg01135.html or
>> http://www.stata.com/statalist/archive/2011-04/msg01348.html or
>> http://www.stata.com/statalist/archive/2009-08/msg01584.html or
>> hundreds of other posts on the topic. In any case, the cluster-robust
>> SE implemented by -xtreg, robust- will deal with serially correlated
>> errors in a more robust way, so I don't see what you hope to gain from
>> -xtregar- (which gives you different estimates, ostensibly more
>> efficient, but tends to perform worse in simulations).
>>
>> webuse grunfeld, clear
>> xtset company  time
>> cap prog drop bsxtar
>> prog bsxtar, eclass
>>  xtset i time
>>  xtregar invest mvalue kstock
>>  end
>> bs, cluster(company ) idcluster(i):bsxtar
>>
>>
>> On Tue, Jan 15, 2013 at 9:45 AM, wasis dat <vasja.sivec@gmail.com> wrote:
>>> Dear Jay V. and Nick C.,
>>>
>>> Thank you for your kind responses!
>>>
>>> I understand that bootstrap doesn't acknowledge the dependence
>>> structure in the panel data. I do not have a clear cluster structure,
>>> just a big panel. The reason why I would still like to use bootstrap
>>> is because my y and x are generated regressors (both T and N are
>>> large), and when y and x are generated regressors they can be
>>> imprecisely estimated. The usual formulas for standard errors do not
>>> account for this. This is why I attempted to bootstrap the standard
>>> errors. Let me say that when ignoring autocorrelation in the residuals
>>> and estimating a FE regression the bootstraped and the calculated
>>> standard errors are practically equal. Of course I have residual
>>> autocorrelation, so I wish to estimate with model with -xtregar. I get
>>> results that are in accordance with my theory, but when presenting a
>>> paper somebody might object that my y and x are generated and so my
>>> standard errors and significance tests are not valid. I wish to avoid
>>> this objection by rather estimating the standard errors using
>>> bootstrap.
>>>
>>> I hope that the above explanation is clear and makes sense. I would be
>>> grateful If you could point me in the right direction (if there is on
>>> of course).
>>>
>>> Kind regards,
>>> Vasja
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/