xtabond2/xtdpdgmm: Estimation of dynamic simultaneous system of equations using GMM

[lakhi narayan]

Hi,
I want to estimate the effect of X on Y in formal sector and then what is effect of Y of the formal sector on the Y of Informal Sector. For instance,
What is the effect of TFP (which is endogenous) on the employment of formal sector and what is the effect of employment of formal sector (due to TFP change) on the employment of informal sector. More specifically,

I have industry-region panel data. Four Time points. Approx 3500 observations. Can I estimate this system of equations simultaneously using GMM by using xtabond2 or xtdpdgmm? Do I need to manually stack the instruments for the equations in order to estimate the system simultaneously using the GMM command? However, I do not think this approach will reveal the mechanism through which Y in the formal sector affects Y in the informal sector. My objective is to estimate the system jointly so that I can identify the transmission mechanism, namely:

TFP in the formal sector → Employment in the formal sector → Employment in the informal sector.

[Manh Hoang Ba]

I think a panel VAR system with 2 endogenous variables (y_formal, y_informal) using the pvar or xtvar commands might be appropriate.

[lakhi narayan]

Thanks, Manh Hoang Ba, for your reply.

I was wondering whether a Panel VAR model would be workable in my case, given that I have only four time periods. Any insights on this would be greatly appreciated. I have two additional questions regarding the use of Panel VAR with my data:

1. If I have four time periods that are equally spaced (e.g., 2000, 2005, 2010, and 2015), can I use a Panel VAR model?

2. If I have more time periods but the intervals are not equally spaced (e.g., 2000, 2005, 2010, 2015, 2021, 2022, 2023, and 2024), can a Panel VAR model still be used in this case?

I would appreciate any guidance on these issues.

[Manh Hoang Ba]

You would not lose additional observations by switching to a panel VAR, provided that you only use VAR(1), since your original specification already includes one lag.
The main issue is data continuity. In both examples you mentioned, using lagged variables in general—not just VAR models—would create difficulties, both in terms of data loss and model suitability.
Even if you declare the data with regularly spaced 5-year gaps:

xtset ivar tvar, delta(5)

I do not think that a value from five years ago (lag 1) would have a meaningful effect on the current period.
Given the context you described, I do not have any additional ideas to suggest.