Query related to two-step system GMM

[Abdullah Ijaz]

I'd like to seek assistance with system GMM. I'm using the following command as an additional robustness check for my fixed-effects regression:

xtabond2 Y L.Y X GDP FDI Manufacturing , gmm(L2.( Y ), collapse) iv(L.X L.GDP L.FDI L.Manufacturing), twostep robust

I have two questions related to this:

1. Are there any issues or concerns with the command I provided?
2. Is the following explanation accurate for the estimation methods used in the command above?

"In order to avoid instrument proliferation and overfitting in my models, I employ the first differences of the second lag of my dependent variable, along with the lagged levels of dependent and explanatory variables as instruments."


Your response shall be appreciated.

[Abdullah Ijaz]

Sebastian Kripfganz Your expert opinion shall be highly appreciated.

[Sebastian Kripfganz]

The iv() option of xtabond2 assumes that all of the variables specified in that option are uncorrelated with the unobserved group-specific effects. This effectively amount to a "random-effects" assumption. This may not be what you want to do.

More on dynamic panel GMM:

• Kripfganz, S. (2019). Generalized method of moments estimation of linear dynamic panel data models. Proceedings of the 2019 London Stata Conference.

[Abdullah Ijaz]

Sebastian Kripfganz Many thanks for your quick repsonse.I apologise for asking such basic questions.

So this statement below would be true in the case of the stata code given below:

Statement: "In order to avoid instrument proliferation and overfitting in my models, I employ the first differences of the second lag of my dependent variable, along with the lagged levels of dependent and explanatory variables as instruments."

Stata code: xtabond2 Y L.Y X GDP FDI Manufacturing , gmm(L2.Y L.X L.GDP L.FDI L.Manufacturing, collapse) iv(i.year, equation (level)), twostep robust

[Sebastian Kripfganz]

Your latest code uses the third lag and further lags of the dependent variable (not its first difference), as well as the second lag and further lags of the explanatory variables as instruments for the first-differenced model. It furthermore uses the first-differenced second lag of the dependent variable and the first-differenced first lag of the independent variables as instruments for the level model. Please see the list of instruments below the regression output.

[Abdullah Ijaz]

Sebastian Kripfganz Thank you for taking the time to provide such a thorough explanation. Your insights have been invaluable in clarifying some of my previous misconceptions. While I found the presentation you shared to be quite technical at times, I am grateful for the opportunity to learn and am confident that with your guidance, I will get there. If you can answer the question below, it will be a huge help for me!



The command: xtabond2 BD BDlagged DT GDPpercapita1 CTP TPN Trade Gov i.Year, gmm(BDlagged DT GDPpercapita1 CTP TPN Trade Gov, collapse) iv( i.Year, equation(level)) nodiffsargan twostep robust orthogonal small


1) According to the explanation you provided above, Am I right to assume that my code uses Second and further lags of the dependent variable, as well as first lags and further lags of the explanatory variables as instrumements for the first-differenced model.
Also, it uses first-differenced 1st lag of the dependent variable and first-differenced first lag of independent variables for the level model.

2) Is there anything wrong with the model specification such as AR(1), AR(2) or Hansen test values?

[Sebastian Kripfganz]

1) With xtabond2, the lag structure can be very confusing when you use the orthogonal option, because the command internally makes a one-period shift. While the list below the output claims that the first lag of BDlagged was used (i.e. the second lag of the dependent variable), it is actually the first lag that was used. If you find this confusing, I recommend to use my [FONT=courier new]https://www.statalist.org/forums/forum/general-stata-discussion/general/1395858-xtdpdgmm-new-stata-command-for-efficient-gmm-estimation-of-linear-dynamic-panel-models-with-nonlinear-moment-conditions]xtdpdgmm command instead. In any case, you should be familiar with the underlying econometric theory. Using the first lag of the dependent variable is actually correct for the orthogonal transformation, while you need to start with the second lag for the first-difference model transformation.

2) Hansen test and Arellano-Bond test look okay.