Question on ppml estimator - RESET test interpretation

[Sehee Kim]

Hello everyone,

recently I'm researching on the gravity model using ppml estimator.
Unfortunately I'm facing a trouble with it so I'd like to ask some advice for it.
.
When I first do the ppml estimator, I used log(export) for dependent variable, and add fixed effects and policy variables.
And when I did the RESET test, most of the results' p-values were bigger than 0.05.
So I thought it would be okay to use the model.

However when I found the literature about ppml estimators in gravity model, most of it says that 'do not take a log for dependent variable'.

So I did ppml estimation with 'export' for dependent variable, and add fixed effects and policy variables.
However, the point is that when I did the RESET test, the most of the results' p-values were 0.000, 0.001, 0.0011 and so on, which are very small.

Therefore I have a question that is it okay to take a model, which uses 'export' for dependent variable and have a p-values of 0.001 or 0.0011.

Thank you so much

Best regards,
Sehee

[Joao Santos Silva]

Dear Sehee Kim,

Indeed your dependent variable should not be in logs; that is the main motivation to use PPML. As for the p-values, it all depends on what p-values you are referring to, so please give further details on the model you are estimating.

Best wishes,

Joao

[Sehee Kim]

Dear prof Joao Santos Silva ,
Thank you for the comment.

My code is like this one, and ccum_wc_1 ccum_shr_1 ccumk_wc_1, ccumk_shr_1 are policy variables.

ppml export YEAR_FE* MONTH_FE* COUNTRY_FE* ccum_wc_1 ccum_shr_1 ccumk_wc_1 ccumk_shr_1 rta

predict fit, xb
generate fit2 = fit^2
.
qui ppml export YEAR_FE* MONTH_FE* COUNTRY_FE* ccum_wc_1 ccum_shr_1 ccumk_wc_1 ccumk_shr_1 rta fit2
WARNING: export has very large values, consider rescaling
WARNING: fit2 has very large values, consider rescaling or recentering
Number of regressors excluded to ensure that the estimates exist: 1
Number of observations excluded: 0

. test fit2 = 0

( 1) fit2 = 0

chi2( 1) = 9.11
Prob > chi2 = 0.0025


The problem is when I did with other policy variables it also shows small numbers of Prob>chi2, such as 0.0000 or 0.0001.

I also consider whether I have to change the model to DID rather than gravity model if I concentrate on the effects of policy variables .

Thank you so much.

Best regards,
Sehee

[Joao Santos Silva]

Dear Sehee Kim,

I suggest you reconsider the set of fixed effects you are using. In particular, consider using origin-time and destination-time fixed effects; you may also want to use the ppmlhdfe command because it implements the same PPML estimator but it is better at dealing with many fixed effects.

Best wishes,

Joao

[Sehee Kim]

Dear Joao Santos Silva

Thank you so much professor.
I'll consider and modify the model again.

Warm regards,
Sehee

[Sehee Kim]

Dear professor Joao Santos Silva,

I did PPML estimation again with ppmlhdfe command and new time fixed effects,
and I got the new chi2 p-values like this:

. test fit2 = 0

( 1) fit2 = 0

chi2( 1) = 4.02
Prob > chi2 = 0.0450

The thing is that it is bigger than before, but it is still less than 0.05 or 0.1.

So I have one more question about it that

Does p-values of chi2 needs to be bigger than 0.05 or 0.1 to use this model.

Thank you.

Best regards,
Sehee

[Joao Santos Silva]

It is us to you to decide what is the relevant significance level. If you have a large dataset, it is OK to use 0.01 as the significance level.

[Sehee Kim]

I really appreciate your advice.

Have a great day

Best regards,
Sehee