logit with fixed effect, cross sectional data

[Abdramane Cherif]

hello, I need your help. I want to estimate a logit on cross-sectional data. in addition, I also want to control the effects of countries, industries and years.

I would like to know how to have similar results

Country FE YES		YES YES			YES
Industry FE	YES		YES	YES		YES
Year FE	YES		YES	YES		YES
Cluster country	YES		YES	YES		YES
Method	LOGIT		LOGIT	LOGIT		LOGIT

Your suggestions are very much appreciated.

Thanks

[Clyde Schechter]

Have you tried

xtset country
xtlogit outcome_var i.industry i.year, fe vce(cluster country)

[Abdramane Cherif]

Thank you so much.

I have tried the command and this is what i get : '' vcetype 'cluster' not allowed ''
r(198);

[Abdramane Cherif]

When i replace xtlogit with logit i find this :

Iteration 0: log pseudolikelihood = -2452.6923
Iteration 1: log pseudolikelihood = -2188.2405
Iteration 2: log pseudolikelihood = -2187.3913
Iteration 3: log pseudolikelihood = -2187.3912

Logistic regression Number of obs = 3573
Wald chi2(20) = 1725.79
Prob > chi2 = 0.0000
Log pseudolikelihood = -2187.3912 Pseudo R2 = 0.1082

(Std. Err. adjusted for 24 clusters in codekpays)

	Robust
self_rationing	Coef.	Std. Err.	z	P>z	[95% Conf.	Interval]
corrupt	.2165579	.1106258	1.96	0.050	-.0002648	.4333805
overdraft_facility	-.5900607	.1376564	-4.29	0.000	-.8598623	-.320259
bank_account	.2090347	.1370699	1.53	0.127	-.0596173	.4776867
access_finance	1.160006	.1646246	7.05	0.000	.8373479	1.482664
credit_commercial	.0138696	.0031138	4.45	0.000	.0077666	.0199726
fonds_propre	.0016878	.0017423	0.97	0.333	-.001727	.0051027
log_ventes	-.0396352	.0223901	-1.77	0.077	-.0835189	.0042485
moyenne_en	.4669297	.156668	2.98	0.003	.1598661	.7739933
petite_en	.6708125	.1898767	3.53	0.000	.298661	1.042964
female_manager	.0463935	.1250372	0.37	0.711	-.1986748	.2914618
experience_manager	-.0074564	.0054827	-1.36	0.174	-.0182023	.0032894
1.manufacturing	.2766	.0892249	3.10	0.002	.1017223	.4514777
a14ya
2011	.8276756	.4104214	2.02	0.044	.0232645	1.632087
2012	.306179	.5162335	0.59	0.553	-.7056201	1.317978
2013	.4692419	.4176778	1.12	0.261	-.3493915	1.287875
2014	.5327559	.4047374	1.32	0.188	-.2605147	1.326027
2015	.6714686	.391274	1.72	0.086	-.0954143	1.438352
2016	.6765817	.3626455	1.87	0.062	-.0341905	1.387354
2017	.5783316	.3743454	1.54	0.122	-.1553719	1.312035
2018	.8387466	.6286813	1.33	0.182	-.3934461	2.070939
_cons	-1.031423	.4335584	-2.38	0.017	-1.881182	-.1816645

[Joro Kolev]

If you want to get feedback on some regression results, always show exactly what you typed at Stata, and exactly what Stata returned.

You can also do what you want by manually including the fixed effects with factor variable notation, i.year i.industry i.country.

For some reason unbeknown to me, - xtlogit, fe- indeed does not accept the option for cluster-robust variance. You can get cluster robust variance by bootstrap. Or by switching to -logit- with factor variable notation as in the previous sentence.

[Jeff Wooldridge]

It looks like you have firm-level data. Is that correct? With N > 3,000 and 24 countries, you should just put in the i.country into the logit command. But, as Joro says, the quality of our responses generally increases with more information. You didn't even show the actually state command.

There's a good reason xtlogit, fe does not support clustering, but Stata is uneven in this. It is known that the so-called fixed effects logit estimator is inconsistent in the presence of serial correlation (unlike, say, FE Poisson or FE OLS). So needing to cluster is admitting your estimators are inconsistent. However, Stata allows this with xtprobit, re and xttobit, re and one is making the same admission. I think it should be allowed with the understanding that needing to cluster means you're not using the best estimation method for a binary response.

[Joro Kolev]

What you are saying is very interesting, Professor Wooldridge, I either have forgotten the fact that -xtlogit- is inconsistent with autocorrelation, or I have never known this fact...

But this is very bad news, because we are out of options if we want to "properly account" for the binary dependent variable's nature. There is no fixed effects probit, and the fixed effects logit has this very undesirable lack of robustness to something that is almost guaranteed to exist in panel data.

Then the question is what do we do if we want to do something that "accounts properly" for the binary nature of the dependent variable, and yet we want our procedure to be robust to correlation within panels...

Is just going with the -xtpoisson- a reasonable alternative option here? Nothing in a textbook Poisson process says that in our sample we have to observe values larger than 1. Can we just go with the -xtpoisson- when we have a binary dependent variable?

It looks like you have firm-level data. Is that correct? With N > 3,000 and 24 countries, you should just put in the i.country into the logit command. But, as Joro says, the quality of our responses generally increases with more information. You didn't even show the actually state command.

There's a good reason xtlogit, fe does not support clustering, but Stata is uneven in this. It is known that the so-called fixed effects logit estimator is inconsistent in the presence of serial correlation (unlike, say, FE Poisson or FE OLS). So needing to cluster is admitting your estimators are inconsistent. However, Stata allows this with xtprobit, re and xttobit, re and one is making the same admission. I think it should be allowed with the understanding that needing to cluster means you're not using the best estimation method for a binary response.

[Abdramane Cherif]

It looks like you have firm-level data. Is that correct?

Indeed, this is the data collected at the firm level.

[Abdramane Cherif]

If you want to get feedback on some regression results, always show exactly what you typed at Stata, and exactly what Stata returned.

xi: logit self_rationing corrupt overdraft_facility bank_account access_finance credit_commercial fonds_propre log_ventes moyenne_en petite_en female_manager experience_manager i.manufacturing i.year , vce(cluster country)

[Abdramane Cherif]

---

My database has this structure:

* Example generated by -dataex-. To install: ssc install dataex

input str23 Country byte(access_finance corrupt overdraft_facility bank_account manufacturing)
"Rwanda" 0 0 1 1 1
"Rwanda" 0 0 1 1 0
"Rwanda" 0 0 1 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 1 0 . 1 0
"Rwanda" 0 0 1 1 1
"Rwanda" 0 1 . 1 0
"Rwanda" 1 0 0 0 0
"Rwanda" 1 0 1 1 0
"Rwanda" 0 0 1 1 0
"Rwanda" 0 1 . . 0
"Rwanda" 0 0 1 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 0 0 0
"Rwanda" 0 1 0 1 0
"Rwanda" 0 0 0 1 1
"Rwanda" 0 0 0 0 0
"Rwanda" 1 0 0 1 0
"Rwanda" 0 0 1 1 0
"Rwanda" 1 0 0 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 1 1 1 1 1
"Rwanda" 0 0 0 1 0
"Rwanda" 1 0 1 1 0
"Rwanda" 1 0 1 1 0
"Rwanda" 1 0 1 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 1 0 0
"Rwanda" 0 0 1 1 0
"Rwanda" 1 0 1 1 1
"Rwanda" 1 0 0 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 1 0 0
"Rwanda" 0 0 0 1 1
"Rwanda" 0 0 1 0 0
"Rwanda" 0 0 1 0 0
"Rwanda" 1 0 1 0 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 1 0 0
"Rwanda" 0 0 1 1 0
"Rwanda" 0 0 0 1 0
"Togo" 1 1 0 1 0
"Togo" 0 0 1 1 0
"Togo" 1 0 1 0 0
"Togo" 0 0 0 0 0
"Togo" 1 0 0 0 0
"Togo" 0 0 1 0 0
"Togo" 1 0 0 1 0
"Togo" 0 1 1 1 0
"Togo" 1 0 1 1 0
"Togo" 0 0 1 1 0
"Togo" 1 0 0 0 0
"Senegal" 0 0 1 0 0
"Senegal" 0 0 0 1 0
"Senegal" 0 0 0 1 0
"Senegal" 0 0 0 1 0
"Senegal" 0 0 . 0 0
"Senegal" 1 0 1 0 0
"Senegal" 1 0 1 0 0
"Senegal" 0 0 1 0 0
"Niger" 0 0 0 0 0
"Niger" . 0 0 0 1
"Niger" 0 0 0 1 0
"Niger" 0 0 0 0 0
"Niger" 0 0 0 0 0
"Niger" 0 0 0 1 0
"Niger" 0 0 0 1 0
"Niger" 0 0 1 1 0
"Niger" 1 1 1 1 0
"Niger" 0 0 1 1 0
"Niger" 0 0 1 1 0
"Congo" 0 0 0 1 0
"Congo" 0 0 1 0 0
"Congo" 0 0 1 0 0
"Congo" 1 1 0 0 0
"Congo" 1 0 0 0 0
"Congo" 0 0 0 1 1
"Congo" 0 0 . . 0
"Congo" 0 0 0 0 0
"Congo" 0 0 1 1 1
"Congo" 1 0 0 1 0
"Tchad" 0 . 0 0 1
"Tchad" 1 0 0 . 0
"Tchad" 0 0 1 1 1
"Tchad" 0 0 0 0 0
"Tchad" 0 0 1 0 0
"Tchad" 1 1 1 1 0
"Tchad" 1 0 0 1 0
"Tchad" 0 0 0 1 1
"Gambia" 0 0 1 1 0
"Gambia" 0 0 0 0 0
"Gambia" 0 0 0 1 0
"Gambia" 0 1 1 1 0
"Gambia" 1 1 1 1 0
"Gambia" 0 1 1 1 1
"Keyna" 1 1 1 1 0
"Keyna" 1 1 0 1 0
end
[/CODE]

[Abdramane Cherif]

My database has this structure:

 * Example generated by -dataex-. To install: ssc install dataex

input str23 Country byte(access_finance corrupt overdraft_facility bank_account manufacturing)
"Rwanda" 0 0 1 1 1
"Rwanda" 0 0 1 1 0
"Rwanda" 0 0 1 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 1 0 . 1 0
"Rwanda" 0 0 1 1 1
"Rwanda" 0 1 . 1 0
"Rwanda" 1 0 0 0 0
"Rwanda" 1 0 1 1 0
"Rwanda" 0 0 1 1 0
"Rwanda" 0 1 . . 0
"Rwanda" 0 0 1 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 0 0 0
"Rwanda" 0 1 0 1 0
"Rwanda" 0 0 0 1 1
"Rwanda" 0 0 0 0 0
"Rwanda" 1 0 0 1 0
"Rwanda" 0 0 1 1 0
"Rwanda" 1 0 0 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 1 1 1 1 1
"Rwanda" 0 0 0 1 0
"Rwanda" 1 0 1 1 0
"Rwanda" 1 0 1 1 0
"Rwanda" 1 0 1 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 1 0 0
"Rwanda" 0 0 1 1 0
"Rwanda" 1 0 1 1 1
"Rwanda" 1 0 0 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 1 0 0
"Rwanda" 0 0 0 1 1
"Rwanda" 0 0 1 0 0
"Rwanda" 0 0 1 0 0
"Rwanda" 1 0 1 0 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 0 1 0
"Rwanda" 0 0 1 0 0
"Rwanda" 0 0 1 1 0
"Rwanda" 0 0 0 1 0
"Togo" 1 1 0 1 0
"Togo" 0 0 1 1 0
"Togo" 1 0 1 0 0
"Togo" 0 0 0 0 0
"Togo" 1 0 0 0 0
"Togo" 0 0 1 0 0
"Togo" 1 0 0 1 0
"Togo" 0 1 1 1 0
"Togo" 1 0 1 1 0
"Togo" 0 0 1 1 0
"Togo" 1 0 0 0 0
"Senegal" 0 0 1 0 0
"Senegal" 0 0 0 1 0
"Senegal" 0 0 0 1 0
"Senegal" 0 0 0 1 0
"Senegal" 0 0 . 0 0
"Senegal" 1 0 1 0 0
"Senegal" 1 0 1 0 0
"Senegal" 0 0 1 0 0
"Niger" 0 0 0 0 0
"Niger" . 0 0 0 1
"Niger" 0 0 0 1 0
"Niger" 0 0 0 0 0
"Niger" 0 0 0 0 0
"Niger" 0 0 0 1 0
"Niger" 0 0 0 1 0
"Niger" 0 0 1 1 0
"Niger" 1 1 1 1 0
"Niger" 0 0 1 1 0
"Niger" 0 0 1 1 0
"Congo" 0 0 0 1 0
"Congo" 0 0 1 0 0
"Congo" 0 0 1 0 0
"Congo" 1 1 0 0 0
"Congo" 1 0 0 0 0
"Congo" 0 0 0 1 1
"Congo" 0 0 . . 0
"Congo" 0 0 0 0 0
"Congo" 0 0 1 1 1
"Congo" 1 0 0 1 0
"Tchad" 0 . 0 0 1
"Tchad" 1 0 0 . 0
"Tchad" 0 0 1 1 1
"Tchad" 0 0 0 0 0
"Tchad" 0 0 1 0 0
"Tchad" 1 1 1 1 0
"Tchad" 1 0 0 1 0
"Tchad" 0 0 0 1 1
"Gambia" 0 0 1 1 0
"Gambia" 0 0 0 0 0
"Gambia" 0 0 0 1 0
"Gambia" 0 1 1 1 0
"Gambia" 1 1 1 1 0
"Gambia" 0 1 1 1 1
"Keyna" 1 1 1 1 0
"Keyna" 1 1 0 1 0
end

[Jeff Wooldridge]

Why did you originally xtset your data at the country level? Shouldn't it have been at the firm level? You have a panel of firms, right?

You could use the correlated random effects approach for unbalanced "panels" that I discuss in my 2019 Journal of Econometrics paper. You can include industry and country dummies if you want. This is the answer to Joro's comment on how we don't have anything that can be done. In my view, for this kind of data structure, the CRE probit is our best option.

[Abdramane Cherif]

Why did you originally xtset your data at the country level? Shouldn't it have been at the firm level? You have a panel of firms, right?

You could use the correlated random effects approach for unbalanced "panels" that I discuss in my 2019 Journal of Econometrics paper. You can include industry and country dummies if you want. This is the answer to Joro's comment on how we don't have anything that can be done. In my view, for this kind of data structure, the CRE probit is our best option.

Thank you
I don't know if it's a panel, since I don't have the time dimension. I rather think that it is a cross sectional data, and I have merged several data collections on firms from several countries.
moreover, when I integrate the country and industry effects I have a problem of multicolinearity.