PPML, panel data

[Tereza Tachovska]

Hello everybody,

I am etimating gravity model of trade in my thesis, I have data on 210 countries over 19 year and I am using panel data, I would like to estimate it with ppml estimator but I do not know how to include ppml and panel data together.
More precisely, I would like to know

• how to the fixed and random effects for ppml
• how to apply hausman test on it

Thank you all!
Tereza

[Joao Santos Silva]

Dear Tereza,

You can use -xtpoisson- to estimate these models, but you may have convergence problems because of the type of data you have. With 19 years of data the results with the RE and FE are likely to be very similar and so there is no much point in doing the Hausman test (the relation between the RE and FE estimator in this context is not as in the linear models).

Notice, however, that the use of RE and FE models is not standard in the trade literature; maybe you should discuss this with your supervisor to make sure that these are the models you want to estimate.

All the best,

Joao

[Tereza Tachovska]

Dear Joao, thank you very much!

I have to admit, I did not notice any other one (it might be caused by fact that both panel data and ppml estimator are relatively recent in the topic: gravity model of trade). I thought that using FE (according to the Hausman test and also according to Wooldridge (2008) or Westerlund, J. & F. Wilhelmsson (2009)) and some steps not to remove the time-invariant variables (for example using dummies for years) will work.

Could you please give me some recommendation on more suitable estimator?

Regards,
Tereza

[Joao Santos Silva]

Dear Tereza,

I would say that the standard in the trade literature is to estimate (by ppml) a model with importer and exporter dummies interacted with time dummies. That is, to include a full set of importer and exporter dummies for each year, but not country-pair dummies. Richard Baldwin has written about this.

All the best,

Joao

[Tereza Tachovska]

Dear Joao,
thank you!
Do not you know about some kind of a "guide" for stata for this issue, xtpoisson command in stata does not seem to be very friendly with additional options.

Thank you!

[Joao Santos Silva]

Hi Tereza,

The kind of models that I described above do not have to be estimated using -xtpoisson-, you can use plain -poisson- or -ppml- (available form SSC).

All the best,

Joao

[Tereza Tachovska]

Hello Joao,
but how to account for panel data there?

[Joao Santos Silva]

You can estimate a FE poisson model by running -ppml- and explicitly including all the country-pair dummies.

Joao

[Tereza Tachovska]

I am afraid this will not work in my case, I am interested in export of France only to 209 countries. Or am I wrong?

[Joao Santos Silva]

Well, you should have said that in your first post! In your case, you can just use -xtpoisson- or -xtpqml-, which is available from SSC. You can also do it with -ppml- and include a dummy for each destination, but it would be slower.

All the best,

Joao

[Milad Aminizadeh]

Hi Joao,

I estimated gravity model with PPML (Fixed Effect). I used robust and cluster options.

PPML dep var indep vars, robust cluster(...)

I have two questions:


1. should I check normality test?why?


2. Should I check Heteroscedasticity test?why?

[Joao Santos Silva]

Dear Milad,

1) No, there is no point in checking for normality because the estimator does not assume normality in any form.

2) Again, there is no need for that because the model does not assume homoskedasticity and it is valid with general forms of heteroskedasticity.

All the best,

Joao

[Milad Aminizadeh]

Dear Joao,

Thanks for your reply.

Milad

[joseph dover]

Hi Joao,

I am currently working on a paper and I was wondering, if I use "xtpoisson......robust fe" , is it correct to say that I am using a PPML FE model? The poisson pseudo-maximum likelihood (PPML) technique (Silva & Tenreyro, 2006)? or am I wrong?

Thank you very much!

Best Regards,

Joseph

[Joao Santos Silva]

Dear Joseph,

If you do not believe that your dependent variable really follows a Poisson distribution and you use robust standard errors, then indeed you can say you are using PPML with FE. Furthermore, if your dependent variable is not a count as in the gravity equation, then you can also say you are using the approach we proposed (Santos Silva and Tenreyro, 2006).

Best wishes and thanks,

Joao