PPML combined with Zero-inflated model?

[Esmee de Bruin]

Hi,

I am using a PPML estimation to predict a pseudo-gravity model about the effect of climate change on migration flow. I have a very large database consisting out of migration flows between over 150 countries in the world. Furthermore, I have included different types of FE (origin and destination-time) in my model

The PPML estimation seems very suited for my type of data and gives promising results. As a robustness check I would also like to explore a zero-inflated model, since I expect that the migration flow sometimes is zero due to a different process than explained by the independent variables included in the PPML (for example low income countries might not report their migrants).

I read on the 'Log of Gravity'- page that zero-inflated models do not work very well for gravity models, because the results will vary depending on the scale of the dependent variable. However, I cannot vary the scale of the dependent variable, as it is always a percentage (of migration).

Is it possible to combine the PPML with a zero-inflated model? Or do I have to use the 'normal' zero-inflated model with FE and robust standard errors?

Thank you in advance for helping me!

Kind regards,
Esmee

[Joao Santos Silva]

Dear Esmee de Bruin,

The comment on the 'Log of Gravity' page if for the case of trade data; migration data is different because in general it is a count. You say that in your case it is a percentage and I wonder why that is the case. Anyway, it percentages ZI models will give you different results if you define the dependent variable to be between 0 and 1 and 0 and 100, and this is not desirable.

More seriously, I am not sure that a ZI model will be valid with all the fixed effects because of the incidental parameters problem; are you aware of any results on this?

Finally, in your context a ZI model implies that the migration flows between some countries is always equal to zero, no matter what happens. This sounds like a very unrealistic assumption.

Best wishes,

Joao

[Esmee de Bruin]

Thank you for your fast answer dr. Santos Silva.

I work with a bilateral migration rate between the country-pairs. So this the amount of people that moved to the destination country as a percentage of the total population of the country of origin. This could indeed be re scaled if you put it like that.

I will look into the incidental parameter problem. I was not aware that this could be an issue.

So you suggest only using PPML and not a zero-inflated model as a robustness check?

Kind regards,
Esmee

[Joao Santos Silva]

Dear Esmee,

Yes, that is what I would suggest. It was not by accident that we suggested PPML to estimate gravity equations; PPML has a number of very nice features that are not shared by other models/estimators.

Best wishes,

Joao