Poisson with fixed effects and clustered standard errors convergence problem

[Camilo Umana Dajud]

Dear all,

I am estimating a gravity model using the Poisson pseudo-maximum likelihood estimator. My model includes origin-year, destination-year and country pair fixed effects. I would like to cluster standard errors at the country pair level. I have the following problems:

- if I estimate the model using xtpoisson the estimation does not converge.
- if I estimate the model using ppml the model converges but no standard errors are reported.

I have in my sample about 120 countries and 30 years. Does anyone know if there exists a solution to this problem?

Thank you in advance!

Best regards,

Camilo

[Joao Santos Silva]

Dear Camilo,

From what you describe, you really need to use -ppml- to get convergence.

About the standard errors, the likely reason for your problem is the presence of dummy variables with just one observation equal to 1 (the so-called singletons). You can investigate this by doing a -summarize- of your dummies using the sample used by -ppml-. That is, after estimating, type

su YOUR DUMMIES HERE if e(sample)

and check if there are any variables with mean equal to 1 over the number of observations used.

Another thing to consider is that you are using a lot of "fixed effects", maybe too many. Among other things, this is likely to create very strong collinearity which will make the estimation of the covariance quite challenging.

Best of luck,

Joao

[Camilo Umana Dajud]

Dear Joao,

Thank you very much for your suggestion. I checked and there are no singletons in the sample ppml keeps for estimation. Would you have any other clues about what might be causing the problem (other than the difficulties associated with the estimation of the covariance with so many dummies)?

Many thanks,

Camilo

[Joao Santos Silva]

Dear Camilo,

Please make sure you follow the advice we give here; that is, create all the fixed effects without excluding the base category and include them all in the model; that is, let the estimator choose which variables to drop (therefore, do not use the xi prefix). If that does not work, it may help you start by removing collinear regressors using the _rmcoll before you actually use the estimator.

Best of luck,

Joao

[Mahdi Gholami]

Dear Prof Silva,
I just had a broad question. how can we cluster standard errors in a gravity model?

[Mahdi Gholami]

Dear Camilo,

Please make sure you follow the advice we give here; that is, create all the fixed effects without excluding the base category and include them all in the model; that is, let the estimator choose which variables to drop (therefore, do not use the xi prefix). If that does not work, it may help you start by removing collinear regressors using the _rmcoll before you actually use the estimator.

Best of luck,

Joao

Dear Prof Silva,
I just had a broad question. how can we cluster standard errors in a gravity model?
Best,
Mahdi

[Joao Santos Silva]

Dear Mahdi,

Just use the option "cluster" to cluster by pair (often people cluster by distance which should do the same).

Best regards,

Joao

[Mahdi Gholami]

Dear Mahdi,

Just use the option "cluster" to cluster by pair (often people cluster by distance which should do the same).

Best regards,

Joao

Thank you so much professor,
but i was thinking that i should cluster by country-pairs? is that right?
Best,

[Joao Santos Silva]

Yes, that is right. In principle clustering by distance is the same thing (assuming that no 2 pairs have the same distance).

All the best,

Joao

[Mahdi Gholami]

Yes, that is right. In principle clustering by distance is the same thing (assuming that no 2 pairs have the same distance).

All the best,

Joao

Thanks a lot Professor,
just one another question, how can i test heteroscedasticity in a xtpoisson estimation?
Best,
Mahdi

[Joao Santos Silva]

You can do it but it takes a bit of work. Do you really need to test for hetero? Poisson regression with robust (clustered) standard errors should be pretty robust to heteroskedasticity.

All the best,

Joao

[Mahdi Gholami]

You can do it but it takes a bit of work. Do you really need to test for hetero? Poisson regression with robust (clustered) standard errors should be pretty robust to heteroskedasticity.

All the best,

Joao

Thank you. All right. so, if it's robust, i don't need to test for hetero anymore.
In general, could you please introduce me a simple reference for reading about heteroscedasticity and the was that can i test it in a gravity model.
I'm so sorry for asking too many questions.
Best,
Mahdi

[Joao Santos Silva]

Dear Mahdi,

We discuss those tests here, and give additional references. However, to use these tests with FE estimators you first need to get fitted values and residuals using something of the kind I suggested here (see the code in #2).

All the best,

Joao

[Mahdi Gholami]

Dear Mahdi,

We discuss those tests here, and give additional references. However, to use these tests with FE estimators you first need to get fitted values and residuals using something of the kind I suggested here (see the code in #2).

All the best,

Joao

Sorry again Professor,
I just read your paper, but i'm a bit confused about how can i perform Park and Ramsey test after getting xtpoisson estimations?
Many thanks in advance

[Mahdi Gholami]

Sorry again Professor,
I just read your paper, but i'm a bit confused about how can i perform Park and Ramsey test after getting xtpoisson estimations?
Many thanks in advance

I mean in Stata