PPMLHDFE Gravity Model - effect of BITs on trade - why Stata omits variables?

[Sara Gimenez]

Dear Joao Santos Silva and Sergio Correia,

I am conducting a research which assesses to what extend BITs increase trade in services. I am using panel data with 656012 observations for the period 2005-2021.

Dependent variable: sum_trade. My question here is whether it is OK to add trade imports and trade exports, or whether I should only regress one of the two. Source of the variable: WTO-OECD Balanced Trade in Services Dataset (BaTiS) — BPM6.

Key independent variable: bit (dummy equal to unity for country pairs sharing a BIT or equal to 0 when only one country is a member of a BIT).

Control variables: Source of the variables is CEPII.

• log_distw
• log_gdp_d
• log_gdp_o
• contig
• comlang_off
• col45
• gatt
• fta_wto

I used the following coding:

cap egen imp_year=group(par year)
cap egen exp_year=group(rep year)
cap egen pair=group(rep par)

cap ado uninstall ftools
cap ado uninstall reghdfe
cap ado uninstall ppmlhdfe

xtset pair year

ssc install ftools
ssc install reghdfe
ssc install ppmlhdfe

ftools, compile

generate sum_trade = TRADEIMP + TRADEEXP
generate log_distw=log(distw_harmonic)
generate log_gdp_d=log(gdp_d)
generate log_gdp_o=log(gdp_o)

ppmlhdfe sum_trade bit log_distw log_gdp_d log_gdp_o contig comlang_off col45 gatt oecd eu fta_wto, a(exp_year imp_year pair) cluster(pair)

The problem is that 6 variables are omitted because of collinearity: log_gdp_d log_gdp_o contig comlang_off col45 gatt. Also, P value is quite high, although the model seems to be explanatory because of the high R-square. Why does Stata omit variables? How can I lower the P value? See the results of the regression:

[Sara Gimenez]

Thank you very much for your help.

[Joao Santos Silva]

Dear Sara Gimenez,

Those variables are omitted because they are collinear with the fixed effects. On your other points:

The p-value is high both because the effect is small and that your variable of interest is likely to vary little in your sample. As for using the sum of imports and exports, it is generally preferred to use just exports (which will increase your sample).

Best wishes,

Joao

[Sara Gimenez]

Dear Joao Santos Silva,

I would like to thank you for your fast response. I would just like to add that I am working with bilateral trade flows data.

As follow-up questions:

• Would you say it is a problem that the variables are collinear with the FE?
  • Would you then recommend eliminating the FE?
  • Otherwise, how could I solve the issue?
• I would highly appreciate if you could recommend some literature where exports are solely used (instead of IMP+EXP) when applying PPMHLDFE.
• I ran the regression only with TRADEEXP with the following coding, and the coefficient of the BIT variable is negative, which is precisely the opposite of what I expected to find: ppmlhdfe TRADEEXP bit log_distw log_gdp_d log_gdp_o contig comlang_off col45 gatt oecd eu fta_wto, a(exp_year imp_year pair, savefe) cluster(pair)

Thank you very much.

Best wishes,

Sara

[Sara Gimenez]

Good afternoon,

As an addendum, I just ran the regression without the FE and found that the P-value is very low. What could explain this?

Disclaimer: I am a new Stata user and I was introduced to econometrics 4 months ago. I actually like it a lot and your research is very inspiring Joao Santos Silva

Thank you once again,

Sara

[Joao Santos Silva]

Dear Sara Gimenez,

1 I would use the fixed effects and not worry about the variables being dropped. They are not needed.
2. Most of the literature does that.
3. Your coefficient is very close to zero and has a large standard error, so I am not surprised it is sometimes negative (and I would not rely on the results without the FE)

Best wishes,

Joao