separation option in the ppmlhdfe

[Bingzi Zheng]

Dear All,

I am using ppmlhdfe to estimate the effect of PTAs on the initiation of trade disputes. The dependent variable is the number of trade disputes between countries i and j for a given year, which contains 84% of zeros since there isn't any trade dispute between many country pairs in a year. I have included control variables such as bilateral trade, real exchange rate, using country#year, target country#year, and country-pair fixed effects. The data is unbalanced, N=25,515 and T=40. prohibit, rules, norules are three mutually exclusive dummy variables indicating the restrictiveness of the trade defense rules in the PTA. Below is my code

ppmlhdfe dispute prohibit rules norules lag1_lnrer lag1_lnimp,  a(year#target_code year#ad_code id) cluster(id)    nolog

The issue is when I add the three-way fixed effects, my regression outcome gives "(ReLU separation check: maximum number of iterations reached; aborting)". It also says "note: 1 variable omitted because of collinearity: prohibit". Besides, the number of observations used in the regression dropped to 15,000. But if I insert separation (none) in the code as below

ppmlhdfe dispute prohibit rules norules lag1_lnrer lag1_lnimp,  a(year#target_code year#ad_code id) cluster(id) separation(none)  nolog

The estimation results do not have the above messages, i.e., prohibit var has the coefficient though it is extremely small (7.31e-09) but with a very big z value ( -87.13). More importantly, no more the aforementioned red message. The number of observations used in the regression is 21,818, out of a total of 25,515.

I was wondering what is exactly separation doing and could I drop this option safely?

I am getting quite lost and I appreciate your help with this issue.

[Joao Santos Silva]

Dear Bingzi Zheng,

I strongly advise you not to ignore the problem and to learn more about it, for example here and in the references linked therein. Besides that, maybe you can try the option separation(mu fe).

Best wishes,

Joao

[Bingzi Zheng]

Dear Joao,

Thank you very much for your very helpful reply. I will carefully go through the material.

Dear Bingzi Zheng,

I strongly advise you not to ignore the problem and to learn more about it, for example here and in the references linked therein. Besides that, maybe you can try the option separation(mu fe).

Best wishes,

Joao

[Fred Lee]

How to identify the separated observations and singleon observations so that I can exclude them manully, then the descriptive statistics I output can also exclude these observations?

[Joao Santos Silva]

Dear Fred Lee

I believe e(sample) identifies the observations used. I am not sure if excluding the dropped observations from the descriptive statistics is a good idea, make sure you want to do it.

Best wishes,

Joao