[Meng Zhang]

Hi, Joao,
I am a student from China. I try to find out if China's outwards foreign direct investment follows comparative motives, that's to say the Chinese firms will locate less-skilled industries to less-skill abundance country. My data is a panel data from 2006 to 2014. And the dependent variable is OFDI_ikt, that is, the number of firms that Chinese firms/ individuals invest in industry k (2-digit SIC code, 20 manufacturing industries in total)in country i in year t. I do the commands below:
egen id=group(country code sic2digit)
xtset id year
then I regress my equation with PPML,
xi: ppml ofdi lndist contig lnhgdp lnchngdp countryskill industryskill countryskillindustryskill bits fta cu ruleoflaw economicfreedomindex financedevelopment i.cntry
and then with xtpoisson
xtpoisson ofdi lndist contig lnhgdp lnchngdp countryskill industryskill countryskillindustryskill bits fta cu ruleoflaw economicfreedomindex financedevelopment, fe vce(robust)
and the interaction variable countryskillindustryskill is my key independent variable.
TO my surprise, the results with PPML seems totally different from the results with xtpoisson. IS there anything wrong with the command? Which result should I believe?

Results.xlsx

Thanks
Best Regards,

Meng

[Joao Santos Silva]

Dear Meng,

It looks as if you are using different sets of FE in the two models. When you use -ppml- you include dummies for 66 countries, while with -xtpoisson- you have 393 groups. Once you sort this out and have the right FE in both models the results should be the same.

Finally note that using the xi prefix with -ppml- is generally not a good idea; you should generate the dummies using the -xi- command with the noomit option, and then include them in the estimation.

All the best,

Joao

[Meng Zhang]

Dear Joao,
Thank you for your reply.
But I just copy the command what you did in the paper Santos(2006), because I just started to learn Stata three days before. Then how to write the right commands?
1) how to use ppml when dealing panel data? I just use xi prefix because I thought xi:ppml is for panel data, and ppml is for cross-section data.
2) how to realize host country effects, industry effects and time effects with ppml? Is xi, pre() i.var to generate dummies?

[Joao Santos Silva]

Dear Meng,

I am not sure if I fully understand your question , but you should create all the dummies you need with the -xi- command and then run -ppml- with these dummies and the other regressors.

If you are using -xtpoisson- with the FE option, make sure you use -xtset- first to explicitly get the panel data setting you want. You will not need to include dummies for the groups defined by -xtset- but you will need to include other dummies, for example year dummies if you want to include those.

All the best,

Joao

[Meng Zhang]

Dear Joao,

I just want to make it sure that if I have written the right code.My denpend_var is in country-sector-year form, there ar 66 countries, and 20 manufacuring industries in 2-digit SIC code from 2006-2014. But my care is only about the China's outwards FDI (OFDI) to 66 host countries, and I try to find out the diterminants of China's OFDI.

so I write the stata commands as below.

egen id=group(countrycode sic2digit)

xtset id year

ppml ofdi lndist contig lngdp lnchngdp countryskill industryskill countryskillindustryskill countryrdpeople industry_rdintensity countryrdindustryrd bits fta cu economicfreedomindex financedevelopment, cluster(id)

Is it right? I can not regress with FE, because my key independent_var is time-variant.

[Joao Santos Silva]

Dear Meng,

If your key independent variable is time-varying you should be able to use fixed-effects (if you want to include them). The rest looks OK to me, but it all depends on what you really want to do. By the way, how many observations and how many clusters do you have?

Best regards,

Joao

[Meng Zhang]

Dear Joao,
After exclusion of some observations, there are 11240 observations, and 1300 cluters. For clarity, I attach the result below. Results.xlsx

[Meng Zhang]

Dear Joao,
If you do not mind, I was wondering if I can write you an email to the adress in Santos(2006). I just can not upload the original data here.

Best Regards,

Meng

[Joao Santos Silva]

This sounds about right but you may want to consider other forms of clustering; for example clustering by country only. Also, consider including fixed effects.

Feel free to email me.

All the best,

Joao

[Anton Ivanov]

Dear Joao, may I kindly ask you the following: Is there any rule of thumb for using robust SE's for (panel) Poisson models? In case data is clustered, that would justify using -vce(cluster id)-. But what if the data is single-level -- what is the rationale for robust SE's?

Thank you in advance for response.

[Joao Santos Silva]

Dear Anton,

The usual standard errors for Poisson regression are not valid unless the data are truly form a Poisson distribution. Since we generally do not believe that to be the case, we should always use robust (or clustered) standard errors.

All the best,

Joao

[Gabriel Arrieta]

Good morning Mr. Santos Silva,

I am doing a dynamic-panel data model to represent the effects of FTAs (Trade Creation and Trade Diversion) in Trade Flows (Exports and Imports) for Peru and the 90% of its trading partners (from 1995 to 2015). Im trying to use the PPML estimator, but I have a question about this: In this kind of models, according to the literature, exists some level of endogeneity between the dependent variable (Exports or Imports) and GDP. For this problem, the authors suggests that we have to use some instrumental variables (in this case: population) to solve the problem of endogeneity. So, in the use of PPML, is there any form to introduce an instrumental variable using the stata commands of PPML estimator? Or maybe exists other form to solve the problem of endogeneity in the gravity model?

Thanks very much for the atention, Mr. Santos.

[Joao Santos Silva]

Dear Gabriel,

Please have a look at -ivpoisson-

All the best,

Joao

[Joyce Wu]

Dear Joao,
I'd like to ask you a question about PPML and FE
Since my value data concludes so much zero value,so I have done a regression using FE and PPML ,But the results seems identical,did the dummy i set was wrong?
the command was as followsiddum is country-pair dummy)
PPML:
quietly ppml value ln_pGDP ln_rGDP ln_pGDPper ln_rGDPper ln_dist language STP PEP PRP GC TD CTP RIR PTI TXR yeardum* iddum* if OECDstatus==1
quietly ppml value ln_pGDP ln_rGDP ln_pGDPper ln_rGDPper ln_dist language STP PEP PRP GC TD CTP RIR PTI TXR yeardum* iddum* if OECDstatus==0

FE:
tsset id year
xtreg ln_value ln_pGDP ln_rGDP ln_pGDPper ln_rGDPper ln_dist language yeardum* STAR PERM PROP CRED PROT TAX TRAD COTR BKRU if OECDstatus==0, fe
xtreg ln_value ln_pGDP ln_rGDP ln_pGDPper ln_rGDPper ln_dist language yeardum* STAR PERM PROP CRED PROT TAX TRAD COTR BKRU if OECDstatus==1, fe

Thanks
Best wishes

Joyce

[Joao Santos Silva]

Dear Joyce,

I do not understand how the results can be identical if you are using different estimators (and it looks like even the regressors are not the same). Can you please post some of your results?

Joao