Multinominal logit, IIA

[Elin Vimefall]

Dear all

I try to model the income strategy of the houshold where the houshold can belong to 4 different groups depending on where they get their income from: Only farm, Farm and wage, Farm and business or mixed. Standard in the literature is to use a multinominal logit.
However, I find that my data violates the IIA assumption.
My problem is then what to use instead? (I do not have any alternative specific variables)


Thankful for all suggestions I can get

//Elin Vimefall

[Stephen Jenkins]

Read the help and Manual entry for mprobit, and see if it is suitable.

[daniel klein]

Have tried fitting a multinoinal probit?

Also read Paul Allison's blog enry about the IIA.

Best
Daniel

[Nick Cox]

No idea here, but just to point out that the term is multinomial.

[Elin Vimefall]

Thank you all!
I have read the manuals and previous posts here.
If I understand it correctly the multinomial probit also requires IIA. (Some suggests the alternative specific multinomial probit, but that would require alternative specific variables, which I do not have).
I have thought about a two step approach, where the households first decides wheter to work off farm at all and then decides among the other groups, however I'm not sure about how to do that...
(It seems like IIA is fulfulled if I only use the groups where the haoushold also work off farm).

Sorry for the bad spelling!

Best /Elin

[daniel klein]

As fare as I remember, the IIA is motivated from the odds (ratio) estimated in logit model. There is no such concept in the probit framework, which therefore does not require the IIA. However, the probit is often difficult to estimate.

Best
Daniel

[Elin Vimefall]

If I understand it correctly, using a multinomial probit is not suitable when the data is not IIA, since the multinomial probit requires independent error terms. I think the multinomial probit that often are suggested is the alternative specific multinomial probit which would solve my problems. However: that model require alternative specific explanatory variables, which I do not have in my data.

Best /Elin

[Maarten Buis]

The IIA is not unique to logit models, though I find it easier to think in terms of logit models. The equivalent problem for probit models is the assumption that the correlation between the error terms is 0, which is assumed by mprobit. In the literature multinomial probit is used in a more broader sense, which allows for also estimating the correlations, though if I remember correctly, identification of that model can be problematic.

[Stephen Jenkins]

My apologies if my earlier reference to mprobit confuses matters. I had recalled -- incorrectly -- its treatment of the error structure. I must have been thinking of asmprobit. Whatever, my reading of the situation is exactly as Maarten spells out clearly. IIA is to do with assumptions about error correlations. Have a look at cmp (on SSC). I think it can fit multinomial probit models with correlated error structures. But note well also Maarten's point about difficulties in fitting and identification in practice.

[Elin Vimefall]

Thank you all for your comments! I think confusing asmprobit with the mprobit is common, since I saw this recommendation to use multinomial probit several times before (it took me some time to understand that it was asmprobit they meant).

Best //Elin

[Insa Thiermann]

Dear Elin,

i'm am facing the same problem you described above and wondered what solution you found to estimate your model? I am struggeling because mprobit requires the IIA as you described above and asmprobit requires alternativ specific variables which i don't have.

Best Insa