logistic regression problem with industry control

[Lorenzo Capasso]

Hello everyone.
I run the following logistic regression (cross section) using eststo in Stata 17, and now I have the following result. What does it mean? Without using i.industry_dummy to control for industry effect I did not have any problem. Is it correct to use i. command to control for different industries effect?


PS: To create i.industry_dummy I use the following code:
egen industry_dummy = group(Industry)


eststo: logistic SectionII_DISSENT20 REMCOM_women Board_women REMCOM_chairwoman lnCEO_pay Institutional_
> Ownership Board_Size REMCOM_size Board_Independence REMCOM_independence CEO_duality lnCEO_tenure lnTotal
> _Asset Stock_Return ROA MBV Leverage Price_Volatility i.Industry_dummy
note: 2.Industry_dummy != 0 predicts failure perfectly;
2.Industry_dummy omitted and 1 obs not used.

note: 3.Industry_dummy != 0 predicts failure perfectly;
3.Industry_dummy omitted and 1 obs not used.

note: 4.Industry_dummy != 0 predicts failure perfectly;
4.Industry_dummy omitted and 8 obs not used.

note: 6.Industry_dummy != 0 predicts failure perfectly;
6.Industry_dummy omitted and 4 obs not used.

note: 10.Industry_dummy != 0 predicts failure perfectly;
10.Industry_dummy omitted and 5 obs not used.

note: 11.Industry_dummy != 0 predicts failure perfectly;
11.Industry_dummy omitted and 2 obs not used.

note: 12.Industry_dummy != 0 predicts failure perfectly;
12.Industry_dummy omitted and 4 obs not used.

note: 13.Industry_dummy != 0 predicts failure perfectly;
13.Industry_dummy omitted and 3 obs not used.

note: 14.Industry_dummy != 0 predicts failure perfectly;
14.Industry_dummy omitted and 6 obs not used.

note: 16.Industry_dummy != 0 predicts failure perfectly;
16.Industry_dummy omitted and 1 obs not used.


Logistic regression Number of obs = 69
LR chi2(-1) = 60.54
Prob > chi2 = .
Log likelihood = 0 Pseudo R2 = 1.0000

-----------------------------------------------------------------------------------------
SectionII_DISSENT20 | Odds ratio Std. err. z P>|z| [95% conf. interval]
------------------------+----------------------------------------------------------------
REMCOM_women | 5.0e+144 . . . . .
Board_women | 0 . . . . .
REMCOM_chairwoman | . . . . . .
lnCEO_pay | 1.43e+10 . . . . .
Institutional_Ownership | . . . . . .
Board_Size | 3.54e+14 . . . . .
REMCOM_size | . . . . . .
Board_Independence | . . . . . .
REMCOM_independence | 6.83e-89 . . . . .
CEO_duality | . . . . . .
lnCEO_tenure | 1.36e-35 . . . . .
lnTotal_Asset | 1.69e-33 . . . . .
Stock_Return | 5.9e+149 . . . . .
ROA | 1.8e-200 . . . . .
MBV | 8.09e-14 . . . . .
Leverage | 0 . . . . .
Price_Volatility | 0 . . . . .
|
Industry_dummy |
2 | 1 (empty)
3 | 1 (empty)
4 | 1 (empty)
5 | 0 . . . . .
6 | 1 (empty)
7 | 4.4e-103 . . . . .
8 | 5.4e+182 . . . . .
9 | 5.6e-139 . . . . .
10 | 1 (empty)
11 | 1 (empty)
12 | 1 (empty)
13 | 1 (empty)
14 | 1 (empty)
15 | 0 . . . . .
16 | 1 (empty)
17 | 6.0e-253 . . . . .
|
_cons | 0 . . . . .
-----------------------------------------------------------------------------------------
Note: _cons estimates baseline odds.
Note: 58 failures and 11 successes completely determined.
(est6 stored)



Thank you for your help

[Carlo Lazzaro]

Lorenzo:
whenever an independent variable predicts the outcome perfectly (ie, all 0 or all 1), the -logistic- machinery suffers, as the MLE cannot go on.
To avoid this nuisance, Stata omits the observations affected by perfect prediction and include in the regression those unaffected only.

[Lorenzo Capasso]

Dear Carlo,

Thank you for your reply.

What should I do? I have this putting into the model i.industry_dummy. Is it correct to use i.x ?

Thank you

[Carlo Lazzaro]

Lorenzo:
plugging -i.industry- in the righ-hand side of your logistic regression is correct (and it s not the issue here).
What bites here is the perfect prediction that affects part of the outcome (i.e., the regressand) when you include -i.industry-.
I would consider getting rid of -i.industry-, start it all over again adding one predictor at a time and see when Stata starts to throw the error message concerning perfect prediction.

[Jackson Monroe]

I would consider using the bayes prefix with some custom priors to regularize the coefficients on your perfectly predictive regressors. The irony of predictors perfectly predicting your outcome is that the only way to run a vanilla logit/probit is to remove them, which is not ideal because they predict your outcome so well. LASSO methods (ridge regression, too) will also fix the problem, and are conceptually similar to bayesian regularization.