[Tobias Wendler]

Thank you so much for condering these issues, I will read further on these issues and report here as soon as I found a satisfying solution

[Tobias Wendler]

Hey Sebastian,

I have a clarification question in the context for predetermined variables.
I assume my Innovation variable to be potentially dependent on shocks in Environmental Impact, but with quite a time-lag (e.g. 3 periods) because Innovation needs time. Also I include Innovation with one-lag anyhow. Thus, I would assume that Innovation is not correlated with the contemporaneous differenced error term.

Is it, due to this argumentation, sound to rely on Innovation as an exogenous variable ?

Thanks so much and best
Tobi

[Tobias Wendler]

Dear Sebastian,

thank you for all the great and very helpful advice provided so far, I hope to catch you with one more question.
I have one more question concerning how GMM instruments. When we allow for example lags 1 to 3. Does the estimator choice the best of these as lags, or does he average them in some way. More specifically this referes to whether if I instrument a variable L0 with lag 1 to 3 and the variable as L1 with lag 0 to 2 would I basically do the same thing ?

Thanks and best wishes
Tobi

[Sebastian Kripfganz]

Both specifications should yield the same results. The two-step GMM estimator (in contrast to the one-step estimator) uses a weighting matrix that "optimally" weighs the moment conditions.

[Tobias Wendler]

Dear Sebastian,
thank you for the help, that was very relevant to know.
When now playing around with time-restrictions I stumbled across one more question of implementation.

I have data for my variable from 1990-2012 and use forward orthogonal deviations.
For the full sample I include yearly time-dummies from 1991-2011.
Concerning some time restrictions I am inconclusive which time-dummies to include, and would like to ask your advice on this:

1) Restriction Year>1997 (starting in 1998): Should one use from 1998-2011 or 1999-2012 ?
2) Restriction Year<2008 (last year 2007): Should one use from 1991-2007 ?
3) Restriction excl. years 2007 and 2008: Should one exclude the dummies for 2007 and 08 ?

I am a little confuse due to the fact that when I include all time-dummies irrespective or excluding specific years the model calculates coefficients for these time-dummies.
If you could provide a short answer on which implementation is correct, or whether several options are acceptable that would be a great help.

Thanks so much and best wishes
Tobi

[Sebastian Kripfganz]

In principle, it should not matter which time dummies you include as long as you include T-1 one of them. That you are using forward-orthogonal deviations does not matter if you are specifying the time dummies for the level model (which I personally recommend even if all other instruments are specified for a transformed model).

If you are excluding certain years from the estimation sample, then you should not specify time dummies for these years either.

I am a little confuse due to the fact that when I include all time-dummies irrespective or excluding specific years the model calculates coefficients for these time-dummies.

Without seeing a small example, I cannot really say what's going on.

[Tobias Wendler]

Dear Sebastian,

thank you for your answer and support. Indeed, I guess I did as you suggested. Using the simple example of wanting to drop the years of financial crises 2008 and 09 from estimation is good to visualize my issue.

First
When typing the specification, excluding t08 and t09 from both the regressors and instruments:

xtabond2 DepVar l.DepVar Explanatory t91 t92 t93 t94 t95 t96 t97 t98 t99 t00 t01 t02 t03 t04 t05 t06 t07 t10 t11 if year!=2008 & year!=2009, gmm(DepVar, lag(2 3) collapse) gmm(Explanatory, lag(2 3) collapse) iv(t91 t92 t93 t94 t95 t96 t97 t98 t99 t00 t01 t02 t03 t04 t05 t06 t07 t10 t11) small nodiffsargan robust nolevel orthogonal

When typing this it gives me the time-dummies except 08 and 09, leaving me with 19 estimated time-dummies, which is equivalent to the amount of maximum observations per cluster.

Second

However, when not explicitly excluding t08 and t09 and typing:

xtabond2 DepVar l.DepVar Explanatory t91 t92 t93 t94 t95 t96 t97 t98 t99 t00 t01 t02 t03 t04 t05 t06 t07 t08 t09 t10 t11 if year!=2008 & year!=2009, gmm(DepVar, lag(2 3) collapse) gmm(Explanatory, lag(2 3) collapse) iv(t91 t92 t93 t94 t95 t96 t97 t98 t99 t00 t01 t02 t03 t04 t05 t06 t07 t08 t09t10 t11) small nodiffsargan robust nolevel orthogonal

the time-dummy for t08 is estimated and only t09 left out, which basically leads to 20 estimated time-dummies although max of obs still is 19 per cluster. Further, the Sargan test indicates in both cases the same number of overidentifying restrictions, although I would expect these to be one lower if there is one more time dummy estimated.

The number of observations remains identical in both specifications, the results qualitatively also.

Basically, I am puzzled by why 20 dummies are estimated (and they are not omitted) and why the degrees of freedom do not differ.

Thanks and best wishes
Tobi

[Sebastian Kripfganz]

The problem lies in the way how the forward-orthogonal deviations are implemented in xtabond2. The instruments are not constructed in the way you think they are. For that reason, I would advise against using forward-orthogonal deviations with xtabond2. You can instead use my xtdpdgmm command. I describe this problem in post #27 of the Statalist topic on the xtdpdgmm command.

[Daniel Toebelmann]

Dear members,

I am currently working with xtdpdgmm and xtabond2 as well. Furthermore, I am also using FOD. My question is a rather formal one. When looking at the following formulation:

Does this formulation still hold, when using FOD? I am wondering, because the delta is used to mark that those are differenced values. However, I do not find any specific formulation for FOD, not even in the initial papers. In which way does the formal description change, when using FOD instead of FD?

I am looking forward to your thoughts and answers!

Best wishes,
Daniel

[Sebastian Kripfganz]

With FOD, the Delta operator is simply replaced by an equivalent FOD operator. In other words, instead of having all of the variables in your formula in first differences, all of them would be orthogonal deviations from their respective "forward mean". In that regard: Yes, the formulation still holds, just with one little caveat: The first-differenced model is well defined for all time periods besides the first period, while the FOD model is well defined for all periods besides the last period.