JWDID in Gravity Setting

[Ridwan Sheikh]

Dear All!
I am using jwdid to estimate the impact of regional trade agreements (RTA) on bilateral trade flows
(1) In the first case, the outcome (trade-flows) is at aggregated-level and varies at exporter-importer-year dimension. I am running the following code

jwdid trade, ivar(id_ci_cj) tvar(year) gvar(RTA_ij) method(ppmlhdfe) fevar(idt_ci idt_cj)

I am interested to obtain simple overall ATT using cohort-specific weights, therefore i ran the following code after jwdid

estat simple [pweight=w_g], predict(xb)

it produces an error that w_g not found.

(2) In the second case, I keep the data at disaggregated industry level. Here, the outcome varies at exporter-importer-industry-year dimension. Therefore, the panel ID, ivar(id_ci_cj_sk), identifies the trade flows between country i (ci) and country j (cj) in industry k(sk) and the tvar(year) is the time variable. Since being an industry-level panel on trade flows across 147 countries across 1990-2010 period, this involves a data set of 16,038,031 observations. I ran the following code to obtain the ETWFE estimates:

jwdid trade, ivar(id_ci_cj_sk) tvar(year) gvar(RTA_ij) method(ppmlhdfe) fevar(idt_ci idt_cj)

This gives the following error message:

reghdfe_standardize(): 3900 unable to allocate real <tmp> [15619903, 109]
GLM: : init_variables(): - function returned error
<istmt>: - function returned error
r(3900)

109 are total number of coefficients ATT(g,t)]'s estimated in our model.

I would greatly appreciate any help in this regard.


FernandoRios

Thanks and regards,
(Ridwan)

[FernandoRios]

vFirst issue. Seems you run out of memory. you may consider upgrading your Ram

First issue, w_g. If that variable exist in our data, then it may be an issue with your version of Stata. Stata 17 had a kind of bug with weights on Margins

[Ridwan Sheikh]

Thank you FernandoRios

The w_g does not exist in my data. I have created these cohort-specific weights as:

* weight uniformly (as is baseline) by number of observations

gen w_unif = __etr__
 
* weight by cohort

egen temp = sum(w_unif), by(FT_EU)
gen w_g = w_unif/temp
replace w_g = 0 if w_g==.
drop temp w_unif

In my case "year" denotes the time period (1990-2010) over which data on trade flows is observed

* weight by event year

gen w_unif = _etr_
egen temp = sum(w_unif), by(year)
gen w_t = w_unif/temp
replace w_t = 0 if w_t==.
drop temp w_unif

Similarily, to create weights by cohort-year and trade I did as:


* weight by cohort*year

gen w_unif = _etr_
egen temp = sum(w_unif), by(FT_EU year)
gen w_gt = w_unif/temp
replace w_gt = 0 if w_gt==.
drop temp w_unif

In my case "trade" is the variable names for my outcome that denote bilateral trade-flows, so to get trade-share weights, I did:

* weight by trade

gen w_unif = _etr_
egen temp = sum(w_unif), by(trade )
gen w_trade = w_unif/temp
replace w_trade=0 if w_trade==.
drop temp w_unif


I hope I am doing it correct, but I am not sure. I would be thankful if you can have a look at these codes and confirm whether I am creating these weights correctly. That is the last part remining.

Thank you very much

[FernandoRios]

Cant tell. Not sure what is your end goal

[Ridwan Sheikh]

Thanks FernandoRios for replying back.

I am following Nagengast and Yotov(2025) and your own paper, Nagengast, Rios-Avila, and Yotov (2025). My goal is to generate the overall treatment effects using alternative weighting scheme (Cohort, Event Year, Cohort x Year, and Trade) and see whether the results are sensitive to using different weights. I am working on a different intervention and its impacts on trade, the coding example above using FTU_EU as an intervention was just for the illustration.

The jwdid did not create these weights and we have to create one, only then the post-estimation "estat simple [pweight=w_t], predict(xb)" will work. I hope this makes sense.

I would still request you if you can provide some confirmation on whether I am creating the event year weights correctly? that would be of great help to me. I am kind of stuck and unable to proceed in the estimation, all with given deadlines. I hope you understand.

Thanks,
(Ridwan)

[Meng Zhang]

Dear @FernandoRios ,

I am working on a similar setting to your recent work with the jwdid package. Specifically, I follow Nagengast, Rios-Avila, and Yotov (2024) closely, but focus on the effects of FTAs in services. I am currently running the following specification:
jwdid trade_sum, ivar(pair_id) tvar(year) gvar(g) method(ppmlhdfe) fevar(exporter_time importer_time) nolog

where g is the first year for treatment.

However, I noticed that the standard errors in the output are missing. I would greatly appreciate your advice on whether there is something wrong with my code, or what I should check or modify to address this issue.

Thank you very much for your time and for developing such a useful tool.

Best regards,
Meng Zhang

[FernandoRios]

My first guess
you are overflowing
Either fixed effects or covariates