AW: st: AW: xtabond2 - Sargan test and reducing instruments

["Dithmer, Jan" <jdithme@food-econ.uni-kiel.de>]

Hi,

I would suggest you carefully read the Roodman papers again:

Roodman, D. (2009): How to do xtabond2: An introduction to difference and system GMM in Stata. Stata Journal 9(1), 86-136. (or the Working paper...)
Roodman, D. (2009): A Note on the Theme of Too Many Instruments. Oxford Bulletin of Economics and Statistics 71(1), 135-158.

If I am not mistaken, it is explained how many instruments are created in difference and system GMM.

And yes, when the system is overidentified you can test whether some variables might be endogenous if you assume the others are not. You can proceed the way you did, just include the suspect regressor in its own iv() option in xtabond2, then it reports a difference-in-Sargan/Hansen test of whether that instrument is valid, assuming the others are. This is equivalent to testing whether the variable is exogenous. For more on this, you may read about the C test in:

Christopher F Baum & Mark E. Schaffer & Steven Stillman, 2003. "Instrumental variables and GMM: Estimation and testing," Stata Journal, StataCorp LP, vol. 3(1), pages 1-31, March.


Hope this helps...
Jan

-----Ursprüngliche Nachricht-----
Von: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] Im Auftrag von Christian Schroetel
Gesendet: Saturday, September 14, 2013 4:37 PM
An: statalist@hsphsun2.harvard.edu
Betreff: Re: st: AW: xtabond2 - Sargan test and reducing instruments

Hey,

I'm not getting further on that.

I now recognized that I get the Sargan test to reject the null if I perform the regression without the level equation (noleveleq) and with only my lagged dependent in gmmstyle with laglimits(0 0). But then I still have second-order autocorrelation.
Remember, my dependent variable is firm growth. If I add the lagged leverage of the firm (which is a regressor and I actually suspected it to be predetermined) as ivstyle()-variable, I get rejection of the null of both the Sargan- and the AR(2) test. Does that mean, adding that variable as instrumental variable is a good choice and I should keep it? Could I still define it as predetermined in gmmstyle()? But still, when I do it like that, I get nearly non of my regressors significant, all with p-values above 0.5. And when I do the same WITH the level equation (system gmm) I get both second-order autocorrelation and overidentifying restrictions. Would you then advise me to use only first-differenced GMM (not system)?

I'm really desperate at that point which is probably due to my lack of knowledge concerning instrumental variables and the like. But the more I read about it, the more confused I get. So, if noone can help me at that particular problem, could someone please at least give me or point me to a somehow straightforward, understandable, non-technical guide to the Arellano-Bond estimations and the like?

I'd be very grateful for any help as I'm not making any progress right now.

Thanks in advance

Christian

2013/9/13 Christian Schroetel <christian.schroetel@googlemail.com>:
> Hey,
>
> ok, so here you go for a bit more detailed information: My starting 
> sample is from about 3k firms (from around 20 countries, some far 
> better represented than others) and years from 1993 to 2012, adding up 
> to around 47k observations. It's unbalanced though and some variables 
> only contain like 12k observations. So, when I combine them in one 
> regression, I get those 3k observations I talked about earlier. Those 
> are then from 445 firms with 1 to 19 years per firm (avg 6.9 years per 
> firm).
>
> Hope that helps a bit.
>
> 2013/9/13 Dithmer, Jan <jdithme@food-econ.uni-kiel.de>:
>> Hi Christian,
>>
>> I suspect that nobody will be able to make any specific comments on 
>> your question, as the number of instruments depends on the number of time periods you have, and you don't say anything about your sample...
>>
>> Best, Jan
>>
>> -----Ursprüngliche Nachricht-----
>> Von: owner-statalist@hsphsun2.harvard.edu 
>> [mailto:owner-statalist@hsphsun2.harvard.edu] Im Auftrag von 
>> Christian Schroetel
>> Gesendet: Friday, September 13, 2013 9:15 AM
>> An: statalist@hsphsun2.harvard.edu
>> Betreff: st: xtabond2 - Sargan test and reducing instruments
>>
>> Dear Statalist users,
>>
>> I'm trying to use the system GMM estimation on my panel data with firm growth as the dependent variable and 13 explanatory variables. One of the explanatory variables is the lagged dependent variable, so I tried the Arellano-Bond, respectively the augmented versions.
>> I've read the help for xtabond, xtdpdsys and xtabond2 and the paper of Roodman but I still don't completely get how that thing is working, in particular how the number of instruments are created. I actually really only want the t-1 lagged dependent variable plus the 12 other explanatory variables, so I tried the following with xtdpdsys (I made it to transform that into xtabond2 as well getting the same number of instruments, but the command would be too long):
>> - xtdpdsys sgrowth l.slnsales slnage sinternationalsales sleverage 
>> srdintensity spersonalpremium sintangibles stobinsq sclr sroa 
>> scurrentratio scashflowsales, maxldep(1) artests(2) -
>>
>> That creates me 49 instruments at about 3k observations and I get the following sargan test:
>> Sargan test of overidentifying restrictions
>>         H0: overidentifying restrictions are valid
>>
>>         chi2(35)     =  990.1915
>>         Prob > chi2  =    0.0000
>>
>> First of all: Why so many instruments? I know those are mostly coming from the dep. variable, because for each indep. variable I remove I get one istrument less, so it's like 35 instruments only from the dep.
>> variable, why is that?
>>
>> Second: What could be reasons the Sargan test statistics is so "bad".
>> I've seen other with only a bit less instruments but far less observations getting far better Sargan tests. What could I do to solve the problem of overidentifying restrictions? May it just be my explanatory variables are bad?
>>
>> Any help would be appreciated, I'm quite near desperation on that.
>>
>> Thanks in advance.
>>
>> Christian
>> *
>> *   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/
>>
>> *
>> *   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/

*
*   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/

*
*   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/