Two step system GMM

[Mohamed Mahmoud]

Hi Sebastian Kripfganz ,


I used two step system GMM for my main regression analysis , I read "Generalized method of moments estimation of linear dynamic panel-data models" but still struggle to apply this to my model.

Data structure :

- Unbalanced panel data from 2006 to 2020 with 6,120 observations

The used command:

 xtabond2 vardep l.vardep  var1 var2 var3 var4 var5 var6 var7 var8 var9 i.fyear i.ind_group_dum, gmm(l.l.vardep  var1 var2 var3 var4 var5 var6 var7 var8 var9,collapse) iv( var2 var3 var4 var5 var6 var7 var8 var9 i.fyear i.ind_group_dum, equation(level)) nodiffsargan twostep robust orthogonal small

My questions :

1. Is the command is correct because I am struggling with gmm( ) and iv( ) , I mean I can treat the variables as endogenous and exogenous and if it is not correct what is the possible suggestions ? using lags and for what is the appropriate lags?

2. I don't know why the number of observations reduced to half from 6120 to 3328 ? is there any other method to maintain the number of observations ?

3. There is an error message that "2006b.fyear dropped due to collinearity" "2011.fyear dropped due to collinearity" "1b.ind_group_dum dropped due to collinearity", I don't understand the reason for this ?

Could you please provide any suggestions or the correct version that I can use for my model

[Sebastian Kripfganz]

1. Whether your specification is "correct" depends on your assumptions. You could 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/page30]xtdpdgmmfe command (part of the xtdpdgmm package) to specify a model based on assumptions about the exogeneity of the variables (and other assumptions). This will show you the respective xtdpdgmm command line (which is similar but not 1:1 transferable to xtabond2).

2. This question cannot be answered based on the available information. You might have many groups with gaps (missing observations).

3. You cannot estimate a full set of dummy variable coefficients; this is known as the "dummy variable trap". Also, you are losing the first period due to the dynamic nature of the model (lagged dependent variable) and the last period due to the forward-orthogonal observation; this explains the omitted time dummies.