estimate gravity model to calculate potential trade by uisng OLS, REM, FEM, TOBIT and PPLM

[CAO DUC SON]

Hi everyone, I'm new here.
I am working on my Ms.thesis. the topic related to DETERMINANTS AND POTENTIAL TRADE OF 25 BIGGEST COFFEE EXPORTERS. (25 exporters, 196 importers, 2001-2019)I have tried to use many methods to estimate the gravity model and the results as in the table bellow
For the linear estimator: I use the Ftest; Breusch and Pagan Lagrangian; Hausman test =>FEM is appreciated, but the H0 of RESET test is not rejected, so there is heteroskedasticity, the predicted values also not good ( the values of some countries is too big). Then I use the PPML model with 3 fixed effects Time-Exporter; Time-Importer and Time-Exporter-Importer
my question is:
- Can I use the Ramsey test to select the appreciated model? (for estimating potential (GAP between predict and actual)
- Can I use the coefficient to estimate the potential trade for an individual country to its importer? (EX: calculate the gap between prediction and actual of Vietnam to all the importers)
thank you so much

	OLS	FEM	REM	TOBIT	PPML Importer fixed effect	PPML Importer, exporter fixed effect	PPML Importer, exporter, time fixed effect
VARIABLES	ln(EX+0.0001)	ln(EX+0.0001)	ln(EX+0.0001)	ln(EX+1)	EX	EX	EX
log_GDPex	0.0970***	0.322***	0.364***	0.464***	1.621***	1.041***	-0.181
	(0.00920)	(0.0764)	(0.0279)	(0.106)	(0.118)	(0.0790)	(0.119)
log_GDPim	0.0666***	2.516***	0.651***	3.115***	-0.00761	2.273***	0.994***
	(0.00902)	(0.0695)	(0.0282)	(0.101)	(0.00806)	(0.152)	(0.153)
log_prodpercap_ex2	-0.0180***	0.0205***	0.0123***	0.0222***	0.0284***	0.0131***	0.0104***
	(0.00120)	(0.00210)	(0.00184)	(0.00276)	(0.00422)	(0.00205)	(0.00312)
log_prodpercap_im2	-0.150***	-0.0108***	-0.0515***	-0.00881**	-0.0417***	-0.000655	-0.00206
	(0.00134)	(0.00284)	(0.00245)	(0.00380)	(0.00193)	(0.00320)	(0.00331)
log_pop_ex	0.505***	0.854	0.0593	1.683	5.761	4.574*	1.375
	(0.0193)	(1.204)	(0.0668)	(1.833)	(5.197)	(2.386)	(1.802)
log_pop_im	1.056***	-0.765**	0.294***	-0.754	0.738***	2.087**	-1.261**
	(0.0145)	(0.308)	(0.0473)	(0.462)	(0.0177)	(0.951)	(0.574)
log_dist	-2.045***		-2.756***	-1.312***	-0.929***	-0.610***	-0.614***
	(0.0302)		(0.103)	(0.0261)	(0.0395)	(0.0257)	(0.0247)
contig	3.185***		2.271***	1.303***	0.626***	1.023***	1.022***
	(0.159)		(0.566)	(0.106)	(0.0908)	(0.0683)	(0.0655)
comlang_off	0.241***		-0.0647	1.668***	-0.0870	-0.305***	-0.308***
	(0.0594)		(0.210)	(0.0553)	(0.0623)	(0.0578)	(0.0569)
landlock_ex	-0.419***		-1.179***	-3.689	4.584	2.070	-1.540
	(0.0674)		(0.232)	(2.902)	(3.285)	(1.512)	(1.152)
landlock_im	-4.061***		-4.056***	-3.189	-0.873***	-33.51***	-5.646**
	(0.0524)		(0.188)	(2.302)	(0.0610)	(4.098)	(2.774)
both_rta	2.952***	-0.351	0.942***	-0.268	3.644***	0.692**	0.745***
	(0.215)	(0.243)	(0.233)	(0.272)	(0.232)	(0.269)	(0.263)
one_rta	0.943***	-0.262	0.322	-0.194	1.890***	0.343	0.419*
	(0.215)	(0.232)	(0.225)	(0.225)	(0.222)	(0.247)	(0.243)
religion	0.836***		0.182	0.103	-0.568***	-0.395***	-0.390***
	(0.0643)		(0.228)	(0.0716)	(0.0783)	(0.0585)	(0.0566)
Constant	-19.66***	-83.03***	-13.72***	-94.87***	-155.5	-181.5***	-19.17
	(0.503)	(21.24)	(1.494)	(34.96)	(99.45)	(46.55)	(35.27)
Observations	87,839	87,839	87,839	87,839	87,839	87,839	87,839
R-squared	0.341	0.034			0.155	0.668	0.702
RESET p-values	0.0000	0.0000	0.0000	0.0000	0.0000	0.1341	0.0003
Ftest		0.0000
Breusch and Pagan Lagrangian			0.0000
Hausman test		0.0000	0.0000
Number of pairid		4,736	4,736
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

[Joao Santos Silva]

Dear CAO DUC SON

I have replied to this in another thread, but I would like to add that nothing is gained by estimating a model using totally inadequate estimators. Just stick to PPML and choose the correct set of fixed effects. Keep in mind that because you have a non-random sample, your results need to be interpreted with great care.

Best wishes,

Joao

[CAO DUC SON]

Dear Professor @Joao Santos Silva,
Thank you so much for your advice!

Yes, it is not a random sample, that makes me confused also,
The 25 countries account for 91% of the coffee exported value(2019), the partner is to all of the importers, But there are some missing when merging so there are 196 importers for each exporter. So I would like to explain the results are for the top exporting countries not for the general.
Base on some previous studies I also have already used the PPLM(Importer-exporter fixed effect) for estimating the potential for each importer.
I also think that the Ramsey test can't use to compare different estimators but can I show it in my result, and combine with theory to explain the reason for selecting my estimator?

Best regard,
Son

[Joao Santos Silva]

Yes, that sounds OK.

Best wishes,

Joao

[CAO DUC SON]

Thank you professor,
your advice helped me so much!

[Alessandra Patti]

Dear Professor Silva Joao Santos Silva

I am studying the effect of corruption on brain drain per italian provinces from 2010 to 2017. I have already used sys GMM with xtabond2 command, Then, i would try to compare my dynamic panel model with a gravity model (which is used too much in skilled migration literature) for finding evidence if sys GMM is better than gravity model at describing evidences or not; although i have read so much on existing literature, I am a bit confused on what procedure I have to use on STATA, I considered the command xtpoisson since it is generally used for panel data (ad my case). I have a binary dependent variable of brain drain and an indipendent continous variables of corruption and other indipendent variables suchc as demografic values and other for indicating quality of university (uni_size and the standardized value of quality of university) . Hence, what type of xtpoisson command do you suggest to use like FE, RE or PA? How do you suggest to write for the command and to perform a post estimation tests?

example . xtpoisson lbraindrain lcorruption lpop lrgdpc uni_size zquality_univ zhospital_mig, irr vce(robust) ??

thank you in advance for suggestionz and your precious comment.

best regards
Alessandra Patti

[Joao Santos Silva]

Dear Alessandra Patti

From your post, it is not clear what are the models that you want to estimate and how you want to compare them. For example, you say that you have a binary dependent variable but in the xtpoisson command you show us it looks as if the dependent variable is in logs, suggesting it is not binary. Also, in the xtpoisson command, the dependent variable should not be logged.

If you give further details on the models you want to estimate, we may be able to help.

Best wishes,

Joao

[Alessandra Patti]

Joao Santos Silva Dear prof Silva
Thank you for your kind Reply. Hence, I have understood now that I need not the logarithm of indipendent variables. My dependent variable is a dummy and the name lbrain is wrong but the variable itself is not in logarithm, it’s a dummy. I will label the variable brain-drain correctly.
hence, my question was related the stata input that I need to insert for evaluating gravity-regression for defying if corruption affects brain drain via xtpoisson demand. How is performed?

xtpoisson brain drain Corruption pop gdp_pc employment quality_univ hospital Migration, irr fe. ?
What are the main post estimation command to proof significance of coefficient withhin the model?
for e camole: estat gof or lrtest ?

another command that i found to be use for gravity model for panel data is xtdpd. What do you suggest on this latter?

however i could be pleased if you would provide me guidelines or suggested article to be read In Order to implement my regression with xtpoisson command.
thank you in advance and have a great day, stay safe
Alessandra

[Jeff Wooldridge]

Alessandra: If your dependent variable is binary, an exponential mean function doesn't make much sense. In fact, I would just use a linear model. To me, the real issue is controlling for lagged y or not. I wouldn't. I think we do this way too much, and it mutes the effect of the other variables. Unless you're really interested in the "state dependence," I would use a linear FE analysis. You can supplement that with a correlated random effects probit and also conditional logit (often called FE logit). In the probit case, you can estimate magnitudes of effects to be compared with the linear model.

[Alessandra Patti]

Jeff Wooldridge Dear prof Wooldrige
hence do you suggest, for my purpose, to use correlated random effect probit? Thank you for you precious advice, very pleased.
Best regards
Alessandra

[Alessandra Patti]

Jeff Wooldridge is this paper a right guidelines for my case? Link at: https://journals.sagepub.com/doi/pdf...867X1301300105

[Joao Santos Silva]

Dear Alessandra Patti

Jeff has already pointed you in the right direction. The only thing I would like to add is that it would be useful for you to first decide on the model you want to estimate, then find the right estimator, and then the corresponding Stata command; xtpoisson and xtdpd estimate very different models so you have to decide on the model before deciding on the command.

Best wishes,

Joao

[Alessandra Patti]

Joao Santos Silva Dear prof
the point is this: I am undecided whether to use xtdpdml command to estimate whether corruption affect brain drain. Hence, when I have used the command xtabond2 I had continuous dependent variable in logarithm as well as continuous indipendent variables in logarithm terms of corruption etc. If I decide to use the xtdpdml command, can I use the same variables in logarithm? Could you provide me an example on how implement such command on stata?

[Jeff Wooldridge]

If you Google “wooldridge Stata Chicago 2011” you should be directed to slides on how to implement CRE probit. The method for binary response is essentially the same as fractional response. You can use probit rather than glm, but it’s not necessary.

[Jeff Wooldridge]

Double post.