Dear all,
I may incur the wrath of very senior colleagues on topics relating to using system-GMM, but I've come to my wits ends and to be honest, right now I am very, very confused.
I'm using Stata 13 and I actually thought I am good with GMM (yes, with difference-GMM)...but applying sys-GMM to my model is going the wrong way.
Having read materials relating to GMM and particularly David Roodman's paper several times, I've tried to set up my model using the format stated in the paper, but my results are getting awry.
It's either the Hansen test statistic is hitting the 'implausible' 1.000 mark, while the Sargan seems reasonable (0.794) and we're told that the Hansen-J statistic is more robust...so attempt to correct the H-J, got me confused the more!
Please I need assistance...I have an unbalanced panel data set. T=45, N=1620, the variables are lnFDI (dep. var) while the explanatory variables are lngdp, lngdppc, credit, natural resources and trade openness.
Any ideas on how to arrange these in Stata using xtabond2?....I've already used pooled OLS and Fixed Effects estimators I only need the sys-GMM estimation to support my argument...and honestly, I think GMM is really complicated (my opinion, though).
Help, pls?
Ngozi
It´s simple to answer you. The confusion in use is derivated by the human perception. Not all systems have the same types of screens and keyborard so, they can not be interpreted and handled similarly, but which causes more confusion is the perception of lighting and software movement. With a mobile phone, makes the image hard to understand, disorienting the subject. And beyond that situation, the subject to see himself lost further increases the difficulty of using such a system such is the level of stress or panic you may have. This perceptive reaction is different from subject to subject as you can understand. There are few studies on human behavior in these situations in emergency climate that I recommend, to the best, their reading.
I hope useful for you.
Kind regards.
...to be honest Silvio, you lost me with your response. But thanks all the same!
Dear Ngozi,
Which variables are considered to be endogenous and exogenous in your model?
In the specification below all endogenous variables will be included in the gmm bracket and the exogeneous in the iv bracket. About the options: For robust standard errors "robust" ; " two" stands for the computation of the two step estimator.
xi: xtabond2 Y l.Y x1 x2 x3 i.year, gmm(l.Y x1, lag (2 2) collapse) iv(x2 x3 i.year) robust two
If the number of instruments is too high, it is advised to collapse the matrix of instrument with the option under the same name.
With regard to the control variables, there might be correlation between lngdp and lngdppc.
Hope this is of help.
Regards,
Nicolas
Nicolas, I can never thank you enough for this simple explanation. I'll adopt and modify with my model....gracias!
Dear Ngozi,
You are welcome. I hope it was helpful. You can also run different models if in the gmm bracket you change the lag structure. It is also advised to keep the number of lags to a minimum of 2.
Indeed you are right, it is not that easy to implement the system GMM estimator.
Surey you already know about the Working Paper by Roodman. "A Short note on the Theme of Too Many Instruments". It can provide you with an insight about the number of lags and number of instruments.
Saludos!
Dear Nicolas,
I have question regarding the variables. I am confused how to decide which be endogenous and exogenous variable.
I am applying GMM to study the effect of health spending, trade openness and military spending on economic growth. I am getting negative coefficients which are not expected.
Could you please let me know more about endogenous and exogenous variables please .
thanks
s
Bosede Ngozi Adeleye
I have applied a two-step system GMM. All the results are consistent and as expected. However, when I add robust to my command, all the coefficients become insignificant.
No of observations is 995. Is it compulsory to apply robust with two-step commands?
Samuel Danilola I did not get any particular answer. However there are certain assumptions for applying robust in system GMM, which you can read from the following article attached with this message. For instance, robust application is more appropriate in finite samples.
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling