Re: st: xtabond2 and consistency using twostep robust

[Manhal Mohammad Ali <manhal.ali@gmail.com>]

Hi,

two step GMM results in more asymptotic efficient estimates than one
step. The bias in the two step standard errors are corrected by
Windmeijer's (2005) correction procedure. Infact if you read like
Baltagi's (2005) textbook I think chapter 8 or 9 on dynamic panel
models also Hayashi (2010) you will see the construction of two step
GMM is different from one step. Two step uses the the consistent
variance co variance matrix from first step GMM. Since the
construction of two and one step GMM are different (both one and two
step use different weighting matrices) it is likely that estimates
will be different. By the way both one and two step are consistent,
but later is more asymptotically efficient. Earlier researchers used
to make inference from one step standard errors since two step
standard errors were biased downwards. But recently due to
Windmeijer's (2005) correction procedure researchers prefer to use two
step (current number of citations from this paper is 1110). If I
remember correctly one step is also equivalent to 2SLS. Check on
Cameron and Trivedi  (2005) chapter 21 on this.

I hope this helps.

Manhal Ali.

On Sat, Aug 6, 2011 at 3:51 PM, Jennifer Kohn <jkohn@drew.edu> wrote:
>
> Dear Statalist --
>
> I am using xtabond2 to estimate a systems GMM dynamic panel data model with both a lagged dependent variable (health) and pre-determined regressors (relationship categories) with a reasonably long panel (18 waves).  I have read Roodman's papers (both "how to" and "too many instruments") many times.  My understanding is that the "twostep" option should only change the efficiency of the standard error estimates, not their consistency.  However, when I run the same specification with and without the twostep (with robust for the Windmeijer correction) I get substantially different point estimates.  In addition, the standard error estimates are not always larger using twostep (again, my understanding is that standard errors are biased down with many instruments so the twostep should bring them up).  Can anyone suggest a reason why I am getting different coefficient estimates using twostep?
>
> Here is my code:
> without twostep:
> xi: xtabond2 H1 L.H1  Dcohab Ddiv Dnever Dwid male age12 lnhhincome highed norgaf i.wave i.region, artests(4) gmm (L.H1 Dcohab Ddiv Dwid Dnever, lag (3 .)) iv(male lnhhincome age12 norgaf highed i.wave i.region) robust
>
> with twostep
> xi: xtabond2 H1 L.H1  Dcohab Ddiv Dnever Dwid male age12 lnhhincome highed norgaf i.wave i.region, artests(4) gmm (L.H1 Dcohab Ddiv Dwid Dnever, lag (3 .)) iv(male lnhhincome age12 norgaf highed i.wave i.region) twostep robust
>
> This particular specification has lag (3 .) but I have experimented with many different lags including the collapse option and adding the twostep changes the point estimates in all of these specifications.
>
> Thank you for any assistance!
> Jennifer
>
> *
> *   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/