Nested logit: the case of no alternative specific variables

[Gabrielle Wills]

Dear nlogit users

I am considering estimating an nlogit model to predict to what extent individual level and school level characteristics predict movements of school principals. The first stage is that they decide either to move out of their current school or stay, and in the second stage the movers either move within the public school system or leave the system altogether. A multinomial logit approach would have been the most sensible approach for the problem at hand but there is strong evidence from a suest and smhsiao test to reject the IIA assumption when applying an mlogit to three outcomes (staying, moving within the system, leaving the system) scenario.

The constraint I have is that I do not have any alternative specific variables. Can you still run a nlogit? The reading I have done, it seems that the likely answer is no. Or can one just interact your outcome variable with non-alternative specific variables to create a pseudo version of these? Some guidance here would be really appreciated. Specifically given my context would it just be better to run a sequential logit?

Best,
Gabrielle

[Ilse Ruyssen]

Dear Gabrielle,

I am facing a similar problem for the moment. Did you find out how to solve this already? I always get an error message: "no cases remain after removing invalid observations" and I think it is because I'm trying to estimate my model without alternative-specific variables. Sorry I can't help you on this...

Kind regards,
Ilse Ruyssen

[Na Chen]

Hi Gabrielle and Ilse,
I am also running a nested logit model with only individual-specific variables. I have already tried using "nlogitrum" instead of "nlogit", by first creating interactive variables for all the individual-specific variables and alternatives. For example, I have four alternatives (all are dummy variables): car (1/0), SUV(1/0), van(1/0) and pickup truck(1/0), one individual-specific variable is income. Then I create car*income = (1/0)*income, SUV*income, van*income, truck*income. I create three nests: car; SUV&van; truck. If I use SUV or van as the base, then the model code works. However, if I use car as the base, it did not work since car is a degenerate nest. Have you even solved this problem? It would be great that we can share our experience. My email address is [email protected].
Thanks,

-Na

[Guest]

Hi everyone,

I would be very pleased if someone could help me. I am desperately trying to find a model to predict discrete outcomes that involve hierarchically ordered decisions (nested alternatives) or process steps by individual characteristics of the respondents (demographics). I also would like to do nest-by-subnest comparisons with respect to the impact of the independent variables. Thus, I use the same set of independent variables at each of the three levels/transitions. Moreover, I intend to calculate marginal effects as well as changes in compound probabilities.

I thought about running a Nested Logit Model in Stata (nlogit) since I would like to consider that the nested decisions are correlated through an individual. However, nlogit requires a choice-specific data structure that contains for each level at least one alternative-variant variable which differs across different alternatives (e.g., price for each available alternative). In my model, the independent variables do not vary across the alternatives. The underlying data do not contain any within-case variability. Moreover, nlogit does not allow to specify the same independent variable in more than one equation (at more than one process step) because it would cause identification problems. Since my data do not contain repeated observations on an individual (e.g., panel data) and I intend to compare the impact of a various variables at different transitions, nlogit seems to be the wrong solution. Am I right? I also thought about using a Multinomial Logit Model (I even tested for IIA), but I would like the model to reflect the structure of the decision process.

Is there any possibility to run a nested logit, given the specific characteristics of my research (e.g., by the use of alternative-specific dummies)? Is there any alternative hierarchical model than nlogit that allows simultaneous estimation of all nests takes into account that the nests are correlated through an individual? If so, what would be the best means of ensuring – as far as possible – unbiased estimators?

Best

Chris

[Marcos Almeida]

Hello Gabrielle and Chris,

Have you already thought about performing multilevel mixed-effects models, perhaps (or not) adding crossed random effects?

Please take a loook on this (http://www.stata.com/manuals13/me.pdf)

Hopefully that helps!

Best,

Marcos

[Guest]

Hello Marcos,

Thank you very much for helping us! I already thought about performing a multilevel approach. However, I am not familiar with multilevel and I rejected the idea after screening the basics about multilevel. Could you please indicate why and how you think this could work?

Best,

Chris

[Marcos Almeida]

Hello Chris,

Please read this thread (http://www.stata.com/statalist/archi.../msg01139.html).


Best,

Marcos

[Ali Souag]

I Na and Else,
I am using the nlogitrum for estimating the transitions.
I am interesting to the moving from Yt-1 to Yt. On t-1, I have 3 states : Employment, Unemployment and inactive. On t, I have 6 states formal employee, informal employee, formal self-employment, informal self-employment, unemployment and inactive. The problem is that I get the error "at least one alternative of connection is contained in multiple alternatives of type; this is not allowed. "Unbalanced data" error.

I have locket in Stata from and I found that for multi-level models, we should define the alternative variable at each level using the alternative variable one level down in the nesting tree.
Na said that he has used SUV or van as the base, my question is how we can define it ?
My adress mail is [email protected]

Best,
Ali

nlogittree e7 e6 e4

xi: nlogitrum i.e7 i.sexe x2 x3 ........,group(id_individu) nests( e7 e6 e4 )

[Embarika Farouk]

Hello all,

I am doing Nlogit Model with Degenerate Nests; no diarrhea and no treatment; as follow:

Child had Diarrhea
/ / \
No diarrhea No treatment Had treatment
/ \ \
Public hospital Private provider Pharmacy

I need to create a constraint to satisfy the RUM condition, however I am getting error constriants () is not allowed.

Also I am doing Nlogit, but I do not have any alternative specific variables:

constraint 1 [no_tau]_cons=1
constraint 2 [notreat_tau]_cons=1

nlogit d || actions: hw1 hw2 || diarrheatreatment: v133 v715 ywipoorer ywimiddle ywiricher ywirichest v481, noconstant case(chid) constraints(1 2)

My question is Can I conduct Nlogit without having any alternative specific variable? Also What I should do to set the constraints?

Highly appreciate your reply

Best
Beky