Dear all,
I am trying to estimate an interaction term between two continuous variables in a PPML model, namely:
where RD is an index of regulatory distance ranging from 0 to 100, and GNI stands for Gross National income. Also, i stands for exporter, j for importer and t for year. The dependent variable (gvc_total) is export from country i to country j in year t. While we expect RD to have a negative effect on export (hence, a negative coefficient), the opposite is true for the moderator GNIit
The coefficient of the interaction term is the one of the output below:
I would have two questions:
1) Is it correct to create an interaction between a logged (RD) and non-logged (GNI) variable when implementing a PPML estimator?
2) Since GNI is not logged, do I need to transform the coefficient of the interaction in order to interpret it?
Looking forward to your feedback and comments.
Thank you in advance for your time
I am trying to estimate an interaction term between two continuous variables in a PPML model, namely:
Code:
β1 (ln(1+ RDijt-1)×GNIit)
The coefficient of the interaction term is the one of the output below:
Code:
asdoc ppmlhdfe gvc_total c.lag1_lnRD_imputed##c.GNIcap_exp lag1_lntariffs lag1_RT
> A ln_dist $dummy_list if exp!=imp, a(exp_sector_time imp_sector_time) vce(cluster
> pair_sector_id) replace nest label cnames(overall GVC participation) dec(3) add(RE
> SET Test, 0.000, AVE, 0.000) d
(dropped 5066 observations that are either singletons or separated by a fixed effect
> )
HDFE PPML regression No. of obs = 657,981
Absorbing 2 HDFE groups Residual df = 68,928
Statistics robust to heteroskedasticity Wald chi2(13) = 8276.65
Deviance = 15296318.24 Prob > chi2 = 0.0000
Log pseudolikelihood = -8466648.213 Pseudo R2 = 0.9144
Number of clusters (pair_sector_id)= 68,929
(Std. Err. adjusted for 68,929 clusters in pair_sector_id)
-----------------------------------------------------------------------------------
| Robust
gvc_total | Coef. Std. Err. z P>|z| [95% Conf. Interval]
------------------+----------------------------------------------------------------
lag1_lnRD_imputed | -.1589669 .0238908 -6.65 0.000 -.2057921 -.1121417
GNIcap_exp | 0 (omitted)
|
c.ln(1+ RDijt-1)#|
c.GNIit | .0029432 .0005393 5.46 0.000 .0018863 .0040001
|
lag1_lntariffs | -.1669346 .0197142 -8.47 0.000 -.2055738 -.1282954
lag1_RTA | .1113005 .0355856 3.13 0.002 .041554 .1810471
ln_dist | -.8107724 .0216022 -37.53 0.000 -.8531119 -.7684329
contig | .3042479 .0290538 10.47 0.000 .2473036 .3611922
comlang_ethno | .1456843 .0371045 3.93 0.000 .0729608 .2184078
comcol | -.229223 .1156892 -1.98 0.048 -.4559697 -.0024762
comrelig | -.0529728 .0549299 -0.96 0.335 -.1606336 .0546879
col45 | .0751228 .0871865 0.86 0.389 -.0957595 .2460052
comleg_posttrans | .2575868 .0244418 10.54 0.000 .2096817 .3054919
transition_lega~e | -.1042719 .0398606 -2.62 0.009 -.1823973 -.0261465
sibling | -.1060025 .079126 -1.34 0.180 -.2610867 .0490817
_cons | 13.01029 .1793332 72.55 0.000 12.6588 13.36178
-----------------------------------------------------------------------------------
Absorbed degrees of freedom:
---------------------------------------------------------+
Absorbed FE | Categories - Redundant = Num. Coefs |
-----------------+---------------------------------------|
exp_sector_time | 12087 0 12087 |
imp_sector_time | 11655 210 11445 |
---------------------------------------------------------+
1) Is it correct to create an interaction between a logged (RD) and non-logged (GNI) variable when implementing a PPML estimator?
2) Since GNI is not logged, do I need to transform the coefficient of the interaction in order to interpret it?
Looking forward to your feedback and comments.
Thank you in advance for your time
