mlogit with fixed effects

[Katharina Meyer]

Hey everyone,

I want to conduct a multinomial logit regression with fixed effects. Is there a stata command available? I use Stata 12.0.

Thanks in advance!

[daniel klein]

Short answer: no there is not. Klaus Pforr was/is working on such a thing (see here) but to my knowledge the ado is not yet available.

You might want to consider a "hybrid-model", as suggested by Allison (2009), although he recently expressed concerns about this approach with non-linear models.

Best
Daniel


Allison, Paul D. (2009). Fixed Effects Regression Models. Thousand Oaks:SAGE.

[Katharina Meyer]

Thank you, I will consider the hybrid-model.

if I wanted to use the mlogit function anyway, what would be a solution to minimize the biase. to include all observed control variables?

[daniel klein]

The hybrid-model actually uses mlogit, as there is no alternative.

Your question is too general to answer otherwise, but including control variables would probably help. There are more fancy methods of dealing with biased estimates, e.g. instrument variables, but there are many sources of bias as well ...

Best
Daniel

[Nathan E. Fosse]

Hi, as you are most likely interested in the observations within clusters, why not use the "mixed" package for non-linear outcomes? The advantage is of avoiding inflating the standard errors (especially if you you have a huge string of dummies to include). You also needn't specify a reference dummy category for those indicators.

If you really must use the "chain-saw" approach, you could do this (with help of Ben Jann's esttab):

qui tab clustervar, gen(dummies_)
eststo: qui logit y x1 x2 dummies_*
esttab ., drop(dummies_*)

That should give you a logit regrssion with dummies included, providing only the relevant output

[Matthew Bergman]

Probably not the most methodologically advanced, but you could run the fixed effect as an "i." variable. The model usually runs.

[daniel klein]

Probably not the most methodologically advanced, but you could run the fixed effect as an "i." variable. The model usually runs.

This is dangerous. While using the ordinary least squares estimator to estimate such (linear) models works well, the same is not true for maximum-likelihood estimation of non-linear models. The coefficients will in general be inconsistent. See Bill Gould's discussion of the topic. Though it dates backs to 2003, the argument still holds.

Best
Daniel

[Katharina Meyer]

I took a look at the hybrid solution provided by Allison. As I have never done a multivarite analysis before and have never worked with stata before its quite difficult for me to understand. maybe you could help me and walk me through the steps i have to conduct.

Some infomation on my thesis: I want to analyse whether personal factors (PF) are influencing the decision/ preferences of students to take a job in a large company (LC) or a small company (SC) after their graduation.
I have a panel dataset. The survey was conducted 5 times.

regression equation looks like this:

Size of company imaginable after graduation = PF1 + PF2 + PF3 + age + sex

Size of company = 3 categories --> LC, SM, LC&SM
PF = time variant
age, sex = time invariant

This is Allinson's instruction how to conduct a hybrid model. As he didn't include the stata commands it would be very helpful for me if you could add the commands and also check whether I understood his instruction right.

1) first step is to calculate person-specific means and deviations from those means for the time-varying predictors (PFs)
2) write the means onto a new data set, with one record for each person.
3) merge the means with the original data set and calculate deviation scores
4) estimate the cumulative logit model
4) run a random effects model with both the personspecific means and the deviations from those means, along effects of time-invariant variables (sex, age)
5) test whether the coefficients for the deviation variables are the same as the coefficients for the corresponding mean variables.

Thank you!!!

[daniel klein]

Allison (2009) has all the code he uses in the appendix, and p. 103 applies here. You may also find Schunck (2013) useful. Note that you do not want the cumulative logit (ordered logit), but the multinomial logit, as your categories are arguably not ordered in a strict sense.

Look at egen's mean() function, combined with the by prefix to get the panel-specific means. I do not completely follow the merge logit outlined here, but I do not know how your dataset is set up either.

Best
Daniel

Allison, Paul D. (2009). Fixed Effects Regression Models. Thousand Oaks:SAGE.

Schunck, Reinhard. (2013). Within and between estimates in random-effects models: Advantages and drawbacks of correlated random effects and hybrid models. The Stata Journal, 13(1):65-76.

[Katharina Meyer]

Is there an objective reason why stata doesn't support or have the function for fixed effects for mlogits?

[Stephen Jenkins]

Katharina: please reread message the informative message (#2) from Daniel Klein and follow the links he provided. Such a model isn't as straightforward as you apparently believe or hope.

[Nathan E. Fosse]

Have opinions on fixed effects multinomial logit models changed?

I ask this question in light of recent developments, notably, the availability of the femlogit package in the ssc archive, written by Klaus Pforr and published in the Stata Journal, volume 14, number: st0362

[daniel klein]

I guess, you actually answered your question. Klaus Pforr has published his command, which implements a fixed-effects multinomial logit. So anyone can now download and use it. The implementation fully documented, so it can be cited and criticized.

Best
Daniel