Executing xtbond2 command using Innovation Drivers in Developing Countries paper

[Yassmine Hassanin]

I'm running the xtbond2 command using the data from the attached paper entitled " Innovation Drivers in Developing Countries paper".
N= 24
T=6
xtabond2 linn l.linn lins lfdi lent lgdpg lhc i.year, ///
gmm(L2.linn lgdpg lhc) iv(lins lent lfdi i.year) twostep robust small

I need to know what the problem is with the above command, I don't get the same results as the paper.

[Yassmine Hassanin]

Sebastian Kripfganz can you please advise on this as I want to correctly specify my model given

[Sebastian Kripfganz]

There can be many reasons for not being able to replicate the original results. I assume you have not obtained the data directly from the authors but re-collected it yourself. In that case, there could be discrepancies in the data set for various reasons - e.g., data revisions by the data providers. The bigger problem often is that authors unfortunately neither provide replication files nor detailed instructions on how they conducted their estimations. In your example, I could not find detailed information in the original paper about how the authors specified the estimator.

[Yassmine Hassanin]

Thank you so much Prof. @Sebastian Kripfganz for your reply. I'm using the dataset provided by the paper (https://link.springer.com/article/10...iuYRVj7OpWksic). The authors added the following footnote to the regression results:

1. Notes: System GMM method used to estimate the models. Syntax xtabond2 two-step small robust noconstant
2. All the variables considered as endogenous. The dependent variable second lagged value used in GMM style and independent variables used in IV style by their levels
3. Sample period 2011–2016. The only countries which have at least three data, the minimum period to estimate GMM-SYS, have been chosen
4. In the first row, “l(LINN)” means lagged value of innovation

My question is that am I writing the stata command right according to the above information?

[Sebastian Kripfganz]

Thank you for pointing towards the table footnote, which provides at least some information. Unfortunately, it is still insufficient to know exactly what they were doing.

I tried to do a quick replication and failed as well. First of all, the statements "All the variables considered as endogenous" and "independent variables used in IV style" are inconsistent with each other. One cannot put endogenous variables into the IV-style option, unless "endogenous" just refers to the correlation with the group-specific error component, but not the idiosyncratic error component. If that's the case, then I would interpret the notes as follows (for the first column in the regression table):

xtabond2 LINN L.LINN LENT LFDI LHC LGDPg, gmm(LINN, lag(2 2)) iv(LENT LFDI LHC LGDPg, eq(diff) passthru) twostep small robust noconstant

This gives me 12 instruments as reported in the table, but the estimates do not coincide.

There is no indication in the paper that year dummies were employed. The note that option noconstant was used is quite concerning. As a general rule, one should never use this option for the system GMM estimator. It forces the regression line to run through the origin, which can severely bias the slope coefficients if the variables do not have zero mean.

Also, following the same syntax, column 2 should only have 13 instruments when the variable LINS is added. It is unclear, why this suddenly shoots up to 20.