Fixed effects logit postestimation

[Mike Clayton]

Hey Statalist,

I'm currently exploring postestimation options for a fixed effects logit model estimated using xtlogit in Stata 13.1.

Theoretically, my understanding is that to generate predicted values/marginal effects for a fixed effects logit (as distinct from a linear fixed effects model), it is necessary to place additional assumptions on the fixed effect. Is this correct?

My reading of the stata manual is that following xtlogit I can use the margins() command with the pu0 or xb prediction options. I know pu0 assumes the fixed effect equals zero. However, I am unsure how xb treats the unobserved fixed effects. Would anyone be able to shed some light on this?

Also, is anyone aware of any other Stata routines I could use for generating marginal effects after an xtlogit?

Any help with these problems would be very much appreciated.

Thanks in advance!!

[Joao Santos Silva]

Dear Mike,

Perhaps this will help: http://www.stata.com/statalist/archi.../msg00889.html.

All the best,

Joao

[Alexander Wuttke]

Dear Joao,

after reading several threads in the stata forum and consulting the literature (brief mentions of this for example in: Longhi/Nandi 2015: A Practical Guide to Using Panel Data, p. 205), I do understand that estimating marginal effects is not recommended and not possible after -xtlogit, fe-.

Though, I still do not understand

1) why this is the case
2) how to best present the results of a fixed effects logistical regression

regarding 1)
In the Thread you mentioned, you say:

The reason is that the marginal effects depend on the value of the FE, which are not estimated.

Could you elaborate on this? Haven´t the fixed effects been estimated in the conditional ML-Estimation that was executed? I still do not understand why marginal effects can be estimated employing an ordinary logistical regression, but not when calculating fixed effects. regarding 2) So, all we can do is presenting the regression tables and the odds ratios despite all the problems in interpreting the results of logistical regressions? Thanks for any help in advance.

[Alexander Wuttke]

It exceeds my comprehension of statistics, but regarding 1) I can point interested readers to Chernozhukov et al. (2009, doi:10.1920/wp.cem.2009.0509) and Fernandez-Val(2009, doi.org/10.1016/j.jeconom.2009.02.007).

[Joao Santos Silva]

Dear Alexander,

With fixed T the fixed effects cannot be consistently estimated. The conditional logit (aka fixed effects logit) is based on a transformation of the problem that eliminates the fixed effects, much like demeaning does in the linear model. Therefore, the conditional logit does not estimate the fixed effects and that is why partial effects cannot be estimated.

About the interpretation of the results, the Kitazawa paper I mentioned in the other thread shows that we can consistently estimate the average (semi) elasticity of the conditional probability of observing y = 1. So, maybe that is the best way to report the results if you want more than just the coefficients. I may even have the Stata code to do that, but I would have to search for it.

All the best,

Joao

[Alexander Wuttke]

The confusion had arisen because although I had understood the method of fixed effects logistical regression,I was misled by the meaning of the term "fixed effects". Now, I know that the "fixed effects" refer to those stable characteristics of a person, that are part of the stable and unobserved heterogeneity, we wish to eliminate using this method (Alison 2005, p. 5, Fixed Effects Regression Methods for Longitudinal Data using SAS; https://books.google.de/books?id=OIP...page&q&f=false).

I will not use this self-proclaimed "revolutionary" (p. 192) method in my paper. Still, if you happen to stumble upon the Stata Code one day, I and the other readers would certainly be greateful to have it at our disposal.

Thanks for you answer.

[Carlo Lazzaro]

Alexander:
it is fequent to mistake conditional fixed effect in -xtlogit- for unconditional fixed effect (like in -xtreg-). By the way, -xtlogit- entry in Stata 13.1 .pdf manual dedicates some lines to explain why -(unconditional) fe- cannot be supported by -xtlogit-

[Joao Santos Silva]

Dear Alexander,

Good to know that all is clear. I have found the code; if you (or any other users) are interested, please send me an email and I'll send you the .ado file. Eventually I would like to write something about this for the Stata Journal, but meanwhile I am happy to make the code available.

All the best,

Joao

[Joao Santos Silva]

Dear All,

In case you find this old thread and are interested in the code, please see here.

Joao

[Badewa Aderogba]

Hello Santos, Your contributions really helped me in my search this morning.
Please, can I have the ado.file. I already sent you a mail requesting for this. Thanks in anticipation of your consideration.

[Joao Santos Silva]

Dear Badewa,

Please see the link in #2.

Best wishes,

Joao

[Gopal Trital]

Dear Joas Santos Silva,

Could you please help me understand how could I possibly interpret the coefficients of an interaction term in fixed-effects xtpoisson model offseted by the variable "population" ?

Model:

xtpoisson count_case policy days_since_incidence policy*days_since_incidence policy*x1_Index policy*x2_index, fe exposure(population)

//sample: 23 countries

//count_case: count of cases [had excess 0s as well, but did not use xtnbreg and zinb models]

//days_since_incidence: is a variable denoting days since the incidence --- this is not the time variable for xtset
//x1_Index and x2_Index are time-invariant continuous variables

policy: categorical variable with three levels: // 0: no policy 1: medium_intensive 2: highly_intensive



Hypothetical results

(beta coefficients)
count_case 11*policy 2*days_since_incidence 0.3*policy*days_since_incidence (-5)*policy*x1_Index (-4)*policy*x2_index

I would be grateful to you for your guidance.

Thank you.

[Joao Santos Silva]

I replied to this here:

Dear Gopal Trital,

Please show us the actual estimation results. Also, note that you should not assume that the policy variable has a linear effect; it would be better to transform the variable into a set of dummies.

best wishes,

Joao