Question on marginal effects in a fixed effects logit and using aextlogit

[Max Nyberg]

Hi,

I am studying an unbalanced panel of companies with a year-industry fixed effects logit model (5 years, 12 industries). The fixed effects have been modeled using least square dummy variables.

In interpreting the coefficients, I have run into texts arguing against interpreting marginal effects in the presence of fixed effects. Could someone explain in simple terms why e.g. average marginal effects cannot be used? If I have understood correctly, the AME calculation ought to take the LSDV into account for every observation in the sample.

Second, I have also found and tried the aextlogit command by Prof Santos Silva. When I attempt to use robust standard errors, this happens:

. xtset industry
       panel variable:  industry (unbalanced)

. aextlogit Y X1 X2 X3 i.year, vce(robust)
vcetype 'robust' not allowed

What could cause this? Many thanks in advance!

[Richard Williams]

aextlogit calls xtlogit. xtlogit says that, with fe models,

vcetype may be oim, bootstrap, or jackknife

As for margins with xtlogit, fe see Steve Samuels comments in

https://www.statalist.org/forums/for...-after-xtlogit

[Joao Santos Silva]

Dear Max Nyberg,

Following up on Richard Williams helpful advice, I have an updated version of the command based on clogit that solves the problem with the standard errors. Please email me and I'll be happy to send it to you.

More generally, estimating a logit model using dummies for the fixed effects will lead to very strong bias unless your T dimension is very large. So, you should be using xtlogit with the fe option.

The problem with the marginal effect is that, as you say, they depend on the fixed effects but you cannot estimate these unless you have very large T. So, without very large T, we do not have a consistent estimator of the partial effects.

Best wishes,

Joao