Re: st: dynamic panel - bootstrapped Sargan and exclusion tests

[Bernardo Schettini <bernardo.schettini@gmail.com>]

Thank you, Stas.

2011/6/24 Stas Kolenikov <skolenik@gmail.com>:
> I have very little clue about dynamic panels and -xtabond-, but you
> will face at least two difficulties.
>
> 1. You need to bootstrap the whole panels (municipalities). That's a
> methodological complication, but, fortunately, not a computational one
> -- you can do this with -bootstrap, cluster() idcluster()-.
>
> 2. If you want to get the boostrap distribution of a test to
> approximate the critical values or get the p-values, you need to
> bootstrap under the null hypothesis. Resampling your raw data won't do
> the trick, as the data satisfy the alternative rather than the null.
> You would need to either get a wicked whitening (or pinkening)
> transformation of your residuals that would produce the data of the
> structure that satisfies the null (correlated within the panel;
> uncorrelated between panels; having the autoregressive structure of
> the desired lag, but 0 lags after that -- none of that is ensured by
> the real data), or just generate them as normal with a given
> covariance structure. Even if you think you are successful with it,
> you still need to sell your approach to the referees.
>
> On Fri, Jun 24, 2011 at 1:01 PM, Bernardo Schettini
> <bernardo.schettini@gmail.com> wrote:
>> Dear Statalisters,
>>
>> I estimating a dynamic panel model via system GMM and difference GMM.
>> My panel is balanced with N=4156 municipalities and T=9 years of data,
>> NT=41560. I need to start with 3 lags of each variable, including the
>> dependent variable, and then test for lag reduction.
>>
>> xtabond2 y1 l(1/3).(y1 y2 y3), gmm(l.y1 l.y2 l.y3) twostep robust
>> xtabond2 y1 l(1/3).(y1 y2 y3), gmm(l.y1 l.y2 l.y3) twostep robust noleveleq
>>
>> The number of instruments is 99 and 121, respectively. The
>> autocorrelation pattern is exactly what one expects from a dynamic
>> panel model (rejects AR(2) but not AR(1) residuals).
>>
>> It happens the Sargan and Hansen tests always reject the moment
>> conditions in system GMM and difference GMM. Reducing the number of
>> instruments won't validate the instruments, and reported p-values are
>> always 0.00.  After some experimentations, I was led to conclude that
>> either (i) overidentifying restrictions are really not valid, or (ii)
>> asymptotic values are leading to overrejection of a correct null.
>>
>> Under (i), I could always use and exactly identified model and produce
>> 2SLS estimates (using the lagged level as the instrument for the 1st
>> difference). Ok, but how to test for lag reduction? Using Wald-type
>> tests? I am not sure if that is correct. With overidentified models,
>> some papers use difference-in-Sargan tests. The problem is my model is
>> just identified under this 2SLS estimation procedure.
>>
>> Under (ii), one alternative is to use bootstrapped critical values. It
>> happens xtabond2 won't accept "boostrap, reps(#)" (it allows for
>> jacknife, but it takes forever). I saw an old discussion in this
>> forum, and some of you had a similar problem. I wonder if anyone can
>> help me create bootstraped critical values for the Sargan test.
>>
>> Thank you,
>> Bernardo Schettini
>>
>> *
>> *   For searches and help try:
>> *   http://www.stata.com/help.cgi?search
>> *   http://www.stata.com/support/statalist/faq
>> *   http://www.ats.ucla.edu/stat/stata/
>>
>
>
>
> --
> Stas Kolenikov, also found at http://stas.kolenikov.name
> Small print: I use this email account for mailing lists only.
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
>

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/