High Estimates When Assessing Currency Union Impact On Trade Using Gravity Equation & OLS/PPML

[Ria Wilde]

Hi,

I am trying to estimate the effect of EMU, CFA franc zone and ECCU on bilateral trade using gravity equation and both OLS and PPML, but I get too high estimates compared to literature for CFA franc zone and especially ECCU. I use the CEPII gravity dataset and create the EMU dummy, CFA dummy and ECCU dummy; I generate logs of continuous variables, generate product of GDPs, groups for FE and add main trading countries for each CU. For control variables, I use lntradeflow lnprod_gdp lndist com_lang com_border fta_wto comcol comcol45 and col_ever. Following the literature, I use the gravity equation to estimate bilateral trade described by Anderson and Van Wincoop (2003). While the estimates for EMU are relatively similar to the ones in literature, the ones for CFA and ECCU are VERY high. I checked for collinearity, outliers etc. and all seems to be in order. This is what I get for OLS:

CU Log-Linear Regression Result
(1) (2) (3) (4)
VARIABLES OLS Fixed Effects Fixed Effects Fixed Effects
EMU -0.186*** 0.217*** 0.150* -0.0443 (0.0549) (0.0625) (0.0772) (0.0890)
CFA -0.417** -0.00950 0.735*** 0.669*** (0.189) (0.172) (0.151) (0.153)
ECCU 1.707*** 2.865*** 2.469*** 2.110*** (0.178) (0.174) (0.211) (0.197)
log product of GDPs 0.816*** 1.046*** 0.692*** (0.00349) (0.00363) (0.0111)
log distance -1.153*** -1.033*** -1.357*** -1.363*** (0.0176) (0.0147) (0.0170) (0.0158)
EU 0.699*** 0.108* -0.261*** -0.115 (0.0486) (0.0613) (0.0700) (0.0755)
regional trade agreement 0.117*** 0.973*** 0.697*** 0.617*** (0.0342) (0.0325) (0.0289) (0.0312)
Constant -10.67*** -19.74*** -4.751*** 18.93*** (0.189) (0.183) (0.412) (0.139)
Observations 897,591 897,591 897,591 993,346
R-squared 0.504 0.629 0.705 0.733
Time FE NO YES YES NO
Importer FE NO NO YES NO
Exporter FE NO NO YES NO
Time-importer FE NO NO NO YES
Time-exporter FE NO NO NO YES
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Any suggestions would be much appreciated 🙏

[Maxence Morlet]

You may want to check the user-written commands ppmlhdfe (you do have panel data, right?)

Note that Santos Silva and Tenreyro (2006) show that coefficients in a log-linear model are only consistent under homoscedasticity, whereas PPML coefficients are consistent if you correctly specify the mean function.

[Ria Wilde]

Yes, I actually used ppmlhdfe

[Maxence Morlet]

Great, Sergio Correia will be glad to hear it I'm sure

[Dr. Iqra Yaseen]

Sergio Correia the problem i face is that of multi-collinearity. my dataset consists of almost same variables as Ria Wilde. but the problem i face is that while dividing my dataset into two separate datasets, one for developed countries and another for developing countries, my control variables are omitted while using ppmlhdfe command in case of developing countries while in case of developed countries i get correct results and doesn't face any multi-collinearity issue. what could be the reason for this. Thanks.