PPML estimator

[Venelina Sekulova]

Hi all

A classmate of mine and I are writing our master thesis on the effect of exchange rates on migration flow towards and from Germany.

We wanted to run a PPML regression of type:

immigrant flow= constant + real_exchange rate+gdp_Germany+gdp_origin country

the sample period is 1990-2008, the migration flows from 168 countries to Germany
real_exchange rates- the bilateral exchange rates between the German currency and each of the respective 168 countries in the sample

We would like to account for the year and country fixed effects, however, we do not know how to include this in the PPML estimator. How do we create these fixed effects and how would they be included in the regression.

Moreover, is it possible to limit the PPML to only one country by including an "if "function that would restrict it to a specific country? Is it possible to do the same for a specific year?

Also, do you think that the PPML model is suitable for our analysis?

Thank you in advance

[Joao Santos Silva]

Dear Venelina,

You can account for those fixed effects simply by creating the corresponding dummies and including them in the model. Notice that in general this approach is not valid, but it works in the special case of OLS and PPML. You can also consider using xtpoisson.

Yes, you can use the "if" to estimate the model using just a sub-set of your data.

Finally, yes, PPML looks very appropriate for what you are doing.

Best wishes,

Joao

[Venelina Sekulova]

Dear Joao,

Thank you for the quick response

Can you elaborate a bit more on how these dummies should be created and included in the regression? What is the Stata command for creating the country and year fixed effects?

How would the "if" function be included in the Stata command, because we tried different commands but Stata does not seem to accept it?

Thank you in advance.

[Joao Santos Silva]

To create the dummies, please look at the xi and tab commands.

To use if do something like this

ppml y x if z>4, cluster(w)

where y, x, z, and w should be replaced by the relevant names.

Best wishes,

Joao

[Venelina Sekulova]

Dear Joao,

In our analysis, we are using

1.)PPML
2.) xtopisson with the fixed effects estimator ( in order to account for unobserved heterogeneity)
3.) PPML with country_* and year_* fixed effects

Sometimes when we compare each of the coefficients of the PPML with the other two models we can see that the coefficient becomes larger or smaller, positive or negative so the change is not consistent across the different explanatory variables.

What can we include as a potential explanation for this change rather than just writing " when controlling for unobserved heterogeneity the coefficient changes ..."?


Thanks in advance

[Venelina Sekulova]

For example you can see in the case below.

[Joao Santos Silva]

Dear Venelina Sekulova,

I am not surprised by the changes because including different sets of fixed effects dramatically changes the specification. I would think carefully about what is the right set of fixed effects and just report the results for that case.

Best wishes,

Joao

[Venelina Sekulova]

Dear Joao,

Based on the particularities that we mentioned in the beginning of this discussion, what would you say would be the best set of fixed effects?

[Venelina Sekulova]

What would you say would be the cause for , for example:

In the case of GDP per capita (for females) in the general PPML specification the coefficient is significant and negative but once we control for country and year fixed effects it becomes positive and significant.

Or for example having an insignificant variable in the PPML that becomes significant once country and year fixed effects have been controlled?


Thank you in advance.

[Joao Santos Silva]

That depends on the correlations between the variables and and fixed effects. It is difficult to predict what happens when we make such big changes to the model.