GMM estimation with Cross-section SUR

[dnfloro1]

Hi,

I am trying to estimate a macro panel data of T=64 and N=16 using GMM. However, I came across a paper by Bjornstand and Nymoen that used GMM with SUR (PCSE)

http://www.economics-ejournal.org/econo ... /version_1


I found the instructions on how to do this in Eviews, but couldnt figure out how to do in RATS. Abny idea how to code-up SUR errors in GMM estimation??

Thanks!!!!

dnfloro

[TomDoan]

I'm a bit confused about what they are doing in that paper. It's a unbalanced data set with residuals with an MA error process. They say that they do GMM, but don't really indicate how the weighting matrices are computed to take the MA process into account. If they did a suboptimal GMM estimator (allowing only for short-run correlations in the own residual), then did a PCSE standard error correction, there are two problems with that:

1. PCSE's don't allow for serial correlation which is clearly present.
2. PCSE's require a balanced sample.

The following is from the Panel Data e-course. While this is describing OLS, the same basic idea applies to GMM as well. It sounds like they might have computed GMM using weight matrices derived under the assumptions in the "Individual HAC Standard Errors", then applied the "Panel Corrected Standard Errors" to correct the covariance matrix, but the actual assumption that appears to be in play is the "Full Panel HAC Standard Errors".

Panel Data Covmat Corrections.pdf

(49.58 KiB) Downloaded 820 times

[dnfloro1]

Hi Tom,

Truly grateful for this reply. I actually bought the Panel e-course book some few days ago before you replied, and it was a big help.

Now, looking at Chapter 7 on Dynamic Panel Data models, the Panel workbooks mentions that GMM or the Arellano-Bond estimator is really more for small Ts as the bias of 1/T disappears when T is large.

This journal that I posted here also has large T=37, but nevertheless claimed they implemented GMM. Some questions:

1. As I am estimating the determinants of TAYLOR RULE equation, with the policy rate as the dependent variable and the lagged policy rate at the right hand side, I find the justification to go further and write up a GMM code. However, i have a large T=64. Does it make sense to go further to run a difference and system GMM? Or should I stop at running FE with clustered standard errors and a 2SLS Anderson-Hsiao estimator?

2 (optional...you can skip this, Tom). How can i modify Roodman (2006) system GMM in chapter 7.1 example of the Arellano Bond GMM)? Roodman (2006) illustrated system and difference GMM in STATA in this journal. http://www.cgdev.org/files/11619_file_H ... 2_1_06.pdf.

Thank you once again!!

dnfloro1