[Kumuthini Sivathas]

Sorry for the late reply.

1 - xtpqml, xtpoisson and ppml with the right dummies should produce exactly the same estimates.

3 - we have a choice between dropping fixed effects or dropping time-invariant regressors; ppml is dropping fixed effects and keeping time-invariant regressors but the estimates of these coefficients are meaningless and they should notbe includedin theregression. Do not userandom effects in this context.

5 - You must be doing something wrong there.

Best wishes,

Joao

Dear Prof. Joao,


In your reply in post #27, you mentioned that the ‘estimates of the coefficients of time invariant variables are meaningless’

Could you explain why is that?

The following paper estimated the coefficients of distance and dummies using PPML as well!

https://ac.els-cdn.com/S002219961500...c1041fb80eacc4

What does it mean?


Thank you.

Kind regards
Kumuthini

[Joao Santos Silva]

If I am not mistaken, the paper you mention uses origin and destination dummies, not pair-fixed effects. So the coefficients that can be identified are different. More generally, it is not possible to interpret the coefficients on any variable that is perfectly collinear with the particular set of fixed effects that is used. I suggest you check a textbook on how fixed effect work.

Best wishes,

Joao

[Kumuthini Sivathas]

Thank you very much Prof. Joao,


I understand the concepts of fixed effect estimation. But when PPML (it’s also a Fixed effect estimation, but using PML estimator, rather than OLS) results did not drop the time invariant variables, I got carried away. Either it did not give any error messages that these variables are correlated with fixed effects.

I want to know the dynamics of PPML in Panel fixed effect contexts. Do you recommend any textbooks or resources for that?

Another question. I’m studying the Exports of one country to all its trading partners. In this case, Can I use origin and destination dummies, like in the paper I mentioned earlier? How should I write the code for such implementation?


Thanks & regards
Kumuthini

[Joao Santos Silva]

Dear Kumuthini,

I suggest you check the books by Jeff Wooldridge and Cameron and Trivedi. I also suggest that you experiment with linear models, estimate using the xtreg and the reg command adding the dummies for the fixed effects and compare the results.

About your second question, because you just have exports from one country, you can only use destination dummies which in this case are equivalent to the usual pair fixed effects.

Best wishes,

Joao

[Kumuthini Sivathas]

This is noted with many Thanks Prof. Joao.


Kind regards
Kumuthini

[Kumuthini Sivathas]

Dear Prof. Joao,

As you suggested, I did run with linear models such as reg command with pair fixed effects (I used destination dummies). The estimates of the coefficients are like using xtreg command.

1. In both cases, I’m using pair fixed effects, right?

I have seen some papers using ‘time fixed effects’ and ‘country time fixed effects’ .

1. What’s the advantage of those?
2. How can I implement them?

However, I would rely on the estimates of the PPML method.

But I’m interested in seeing the coefficients of time invariant variables. To see that, what if I use a Pooled OLS.

1. What’s your opinion?

I know that I cannot control for omitted variable bias and endogeneity biases, still I can report about the signs, right!!

Kindly share your thoughts please.

Thanks & regards
Kumuthini

[Joao Santos Silva]

Dear Kumuthini,

I am afraid I do not think that there is any reasonable way to obtain the coefficients you want to estimate. I suggest you discuss this problem with your advisor.

Best wishes,

Joao

[Kumuthini Sivathas]

Okay. Noted with many Thanks Prof. Joao. I understand the seriousness of using pooled OLS. I will not use it.

[Kumuthini Sivathas]

Dear Prof. Joao,


I understand that PPML method controls for country pair fixed effect. But i could not figure out whether PPML method also adds the country time fixed effects or do i have to include it externally?

Could you please clarify this?

Thanks & regards
Kumuthini

[Joao Santos Silva]

Dear Kumuthini,

By default, ppml does not control for any fixed effects. You can, of course, use ppml to estimate a model where fixed effects are included as dummies, but these need to be explicitly included in the model by the user.

Best wishes,

Joao

[Kumuthini Sivathas]

Okay. Noted with thanks Prof. Joao. The following code, adds Country pair fixed effects dummies to my model. xi: ppml export fta lnprodgdpc lnprodgdp i.dcode, cluster(dcode) If my time variable is 'Year', Could you please show me how to include Country time fixed effects into the model, along with the above Country pair fixed effects? Thank you very much. Kind regards, Kumuthini

[Kumuthini Sivathas]

Okay. Noted with thanks Prof. Joao.

The following code adds Country pair fixed effects dummies to my model.


xi: ppml export fta lnprodgdpc lnprodgdp i.dcode, cluster(dcode)

If my time variable is 'Year', Could you please show me how to include Country time fixed effects into the model, along with the above Country pair fixed effects?

Thank you very much.

Kind regards,
Kumuthini

[Joao Santos Silva]

Dear Kumuthini,

I would create all the dummies before estimation (using the xi command) and then just estimate the model with all these dummies included. You may want to use ppml_panel_sg to estimate such model as it includes the fixed effects automatically.

Best wishes,

Joao

[Kumuthini Sivathas]

Okay Prof. Joao. Thank you very much.

I will try both the methods.


Kind regards
Kumuthini

[Kumuthini Sivathas]

Dear Prof. Joao,


How can I do a robustness check for ‘PPML’ method? Is it good to check against another methodology? Or do I have to add some control variables and (I’m concerned as there is no better method than PPML in gravity context) discuss the results?


Kindly advise me.

Thanks & regards

Kumuthini