GMM model (large T and small N)

[rozita saadan]

Dear Statalist,

I have panel data with T=50,000 and N=1500. Because the model has lag dependent variable, I would like to use a dynamic model using GMM. However, I understand that the GMM model is efficient for large N and small T.






I have panel data with N=10, T=50. I would like to estimate a dynamic panel but as I understand Arellano/Bond type estimators are for large N, small T. Can I use these estimators for my data or what is the best way to proceed? Thanks a lot for any help.

[Andrew Musau]

These issues are well documented in standard econometric texts. Have a look at this handout and pay special attention to pp.22- 38 where there is a discussion of Arnold Zellner's seemingly unrelated regression estimator (SURE) and p.38 onwards: why static panel models are not appropriate with dynamic panels.

http://fmwww.bc.edu/ec-c/s2013/327/EC327.S2013.nn2.pdf

[rozita saadan]

Dear Andrew Musau,
Thank you so much for the source of material.

[ANTONIO NETO]

Andrew,

I saw your answer and I think you could help me. Let me say about my case:

I´m trying to estimate effciency in brazilian banks using the Almost Ideal Demand System (AIDS) , as proposed by Hughes and Mester (2013).

The model has fIve equations, but one is dropped. So the four equations are: one profit share equation, two input share equations and capital demand.

The demand for capital is derived from a first-order condition for optimal level of equity capital, which, according to the authors, is a conditioning argument in the share equations.


My dataset is a panel data, unbalanced with gaps, quartely observations from 2001 to 2009.

I read various articles in STATA literature and most of them use certain commanands that are appropriate for the estimation of demand of goods.

The system proposed by the authors has, beyond the AIDS´s traditional regressors, includes outputs, risk proxy and equity capital.

Most of the sugestions uses the commands related to SURE, but all applyed to data with a single unity of time (year).

I found the command xtsur, but unfortunately this command didn´t work. The computer has taken 15 hours with no results. I think it´s a problem of convergence.


Sorry for this long report, but I am desperate to solve this.

Do you have some advice for me ?

Do you think that would it be consistent using my panel as a pooled cross-section and apply the sure command?

On the original paper, the authors say that he following: " we estimated te model using nonlinear two-stage least squares".

Again, Thanks for your attention.

Antonio Neto

[Roman Mostazir]

Antonio, it is not a recommended practice to ask a different question into a topic started by somebody else. You need to start a new topic of your own. Also it is not enough just to say something did not work. You need to be more specific by providing Stata (it is not STATA) commands and outputs to let us know what did not work. Please see the FAQ section for posting rules. The section also contains information on how to use code blocks if you are to provide Stata outputs and data examples.

[Andrew Musau]

Hi Antonio, please start a new thread as adviced by Roman and then myself and others will have a look.