[Dirk Foremny]

Jeff Wooldridge, I was going through your lecture notes regarding this last night. Might this case of a Poisson first stage be comparable to the example you give about binary first stages? All the best!

[Dante Donati]

Dear Jeff Wooldridge
The solution you suggested was extremely helpful.

We have read the paper and managed to set up the procedure, but are still a bit unsure about the appropriate implementation of the standard error estimation.
I'd like to ask if you could give more details on how to calculate the errors in the first and second stage or if you have some references to look it up.

Thanks! All the best,
dante

[Jeff Wooldridge]

Sorry for the delayed response, Dirk. I have the Stata forum stuff go to my gmail account but then I forget to check it. You are correct in that you can use an approximate solution with a Poisson first stage. I would use the generalized residual and insert that into the second stage.

Dante: How large is your cross sectional dimension? How large is T? If N is fairly large, the easiest thing is to use the panel bootstrap on both stages. Similar to what is done here: PW (2008).

[Sebastian Kraus]

Dear Jeff Wooldridge, Would you have any guidance on this two stage procedure for a distributed lag model as the Poisson second stage? Thanks so much.

[Dante Donati]

DearJeff Wooldridge , that's extremely helpful. Thank you. We have T=10 and N=8,000. I guess that can be considered a fairly large N.
We'll follow your advice. All the best,

dante

[Jeff Wooldridge]

Sebastian: If you are treating the variable that appears with a lag as enogenous then the CF approach is tricky. It would be better to use the kind of IV approach discussed in the paper by Windmeijer (2000, Economics Letters). He shows how to fix some moment conditions that I had proposed in a 1990 working paper. It's a transformation effectively eliminates the heterogeneity and then you apply GMM to the resulting moments.

Dante: Glad it helped!

[Sebastian Kraus]

Thanks, Jeff Wooldridge, will read carefully.

[Mora Adet]

Dear Jeff Wooldridge . I would be glad to hear your opinion on how to proceed with my estimations. I have a panel containing 80 countries over 23 years, where y is a count variable and I have 5 continuous x's. Two of these x's are endogenous and I wonder if the most appropriate approach to estimating my model is the nonlinear GMM procedure (Mullahy, 1997; Windmeijer, 2000) rather than the control function approach? Your advise would be greatly appreciated!

[Dirk Foremny]

Thanks a lot Jeff Wooldridge, that was really helpful! Very kind of you giving advice here in the forum!

[Filippo Santi]

Dear Professor Jeff Wooldridge ,

Thank you for your post above, I found the approach extremely interesting. I have a question on how to extend the strategy you proposed to a panel dataset of roughly N = 120*120 countries, T = 14. All my specifications include country*time fixed effects for both origin and destination (some specifications also include country pair FE). Given the number of fixed effects involved, I resorted to use the ppmlhdfe command (developed by Tom Zylkin , Sergio Correia and Paulo Guimaraes ) to obtain non-IV estimates, which does not require to declare the panel structure of the dataset (as it pools all observations together).

Since both my endogenous and instrumental variables are continous, I was wandering whether the same approach could be adopted in presence of high dimensional fixed effect (HDFE) that is, replacing the initial equation

 xtreg y2 z1 ... zJ zJp1 ... zM i.year, fe

with something like

reghdfe y2 z1 ... zJ zJp1 ... zM , absorb(origin_time destination_time [country_pair]) vce(cluster country_pair)

Filippo

[Olov Isaksson]

Hi Dante:

Unfortunately, there is no statistical justification for the procedure you propose. It could suffer from the incidental parameters problem, although maybe using fixed effects in the second stage eliminates that. But I do know that to justify plugging in fitted values into an exponential function imposes strong assumptions.

What is the nature of X? Can it be treated as roughly a continuous variable? If so, I have an easy solution for you. If X is discrete, I have a somewhat harder solution for you (but not so hard for someone with decent Stata programming skills).

Best,
Jeff

Dear Jeff,

I would be very curious about this "harder" solution if the EEV is not continuous. I am working on a problem with a binary EEV and directly implementing the method you suggest here (for continuous variables) seem to give strange results. Thanks in advance!

Best regards,
Olov

[Rie Chuang]

Dear Jeff Wooldridge,

Could I please confirm that the control function approach you set out in Lin and Wooldridge (2017) applies to PPML? In the sample code that you provided (pasted below), would it be acceptable to use ppml or ppmlhdfe instead of xtpoisson? Thank you!

 
 xtreg y2 z1 ... zJ zJp1 ... zM i.year, fe predict double v2h_fe, e xtpoisson y1 y2 v2h_fe z1 ... zJ i.year, fe vce(robust)

[Alessandra Patti]

Dear Prof. Jeff Wooldridge
i have one question concerned static panel data model: I would use gravity model and I have a balanced panel dataset of 110 italian provinces and 7 years time period. I have too many variables which are 0 In performing a gravity model, should I include all 0 into regression or do I regret 0s and do not insert them? I was considering to use the 0 inflated poisson regression. What do you think on this latter model that I mentioned?Thank you in advance for your suggestions
best regards
Alessandra

[Joao Santos Silva]

Dear Alessandra Patti,

It looks as if Jeff did not see your question. For what is worth, my view is that you need to keep all the zeros and should not use a zero inflated model; note that the results of a zero-inflated model will depend on the scale of your dependent variable, and that is not a good thing...

Best wishes,

Joao

[Jeff Wooldridge]

I did miss a few questions here. I again agree with Joao: just use the Poisson FE estimator and keep the zeros.

For Rie: ppml and ppmlhdfe can be used with the control function approach, but the former will not remove fixed effects unless you include cross-sectional dummies. If in ppmlhdfe you absorb the cross-sectional identifier (and the time identifier) you should get the same as xtpoisson, fe when you also include time dummies.