Alternative-specific coefficients and attributes in multinomial logit

[Alessandra Motz]

Good morning,


My name is Alessandra, I am a new Stata user.

I am analysing data from a discrete choice experiment, and I have some problems in implementing the model I have in mind.
I'd like to start with a simple multinomial logit, but with:
- alternative-specific coefficients for some of the alternatives' attributes (e.g. the price of the alternatives, so that I can account for the fact that consumers have different sensitivities to price variations in each alternative)
- an attribute that characterizes only one of the available alternatives, and is not present for all other alternatives.

In short, the model I have in mind is something like this:

V1 = alpha1 + beta_price1 * price1 + beta_quality * quality1 + beta_greencontent * greencontent1 + epsilon1
V2 = alpha2 + beta_price2 * price2 + beta _quality * quality2 + epsilon2
V3 = beta_price3 * price3 + beta _quality * quality3 + epsilon3

I have tried with asclogit and mixlogit, but when I try to include my "greencontent" attribute among the indepvars in the asclogit command, I get the following error message:

no cases remain after removing invalid observations

So I guess that Stata cannot estimate the multinomial logit if there are missing values for "greencontent" in all alternatives but one for every choice task.

Moreover, I can't find any option for allowing for alternative-specific coefficients for my price variable.

I'd be really grateful for any suggestion regarding options or alternative commands to do this.
Thanks in advance,
Regards,


Alessandra

[Doug Hemken]

Not sure I understand. Green content is 1 for V1 and 0 for V2 and V3? or missing?

[Alessandra Motz]

Greencontent ranges from 40% to 100% for V1, and is missing for V2 and V3. Maybe I should add that my alternatives are labelled.

[Alessandra Motz]

Just a short update in case someone should face the same problem.

I didn't find out how to do this with Stata so far. Hence, I reverted to Biogeme, that allows me to specify the exact shape of each alternative's utility function - but unluckily requires a lot of effort for postestimation analyses.

If anyone should have any hint on how to proceed with Stata, I'd be happy to read it!

[Floris Dalemans]

Hi everyone,

I am also analyzing data for a labeled DCE and also facing 1 of the 2 mentioned problems: getting alternative-specific coefficients for the attributes. In my case, each choice set contains 3 labeled alternatives and an opt-out, leading to following utility equations:

So, through the command asclogit Stata easily allows for estimation of alternative-specific constants (= beta01, beta02 and beta04), but not for estimation of alternative-specific attribute coefficients (beta11, beta12 and beta13 cannot be estimated; only a single coefficient for attribute 1 is estimated - similar for beta21, beta 22, beta23).

Maybe by now anyone knows a workaround for this in Stata? After this conditional logit, I also want to estimate a latent class model (command lclogit) for which I will probably run into the same problem; maybe there exists a similar solution.

Alessandra, Nlogit is another software allowing explicit definition of utility equations. However, to the best of my knowledge it is rather limited in accounting for attribute non-attendance; therefore I decided to try out Stata (which has for instance the eaalogit command).

Many thanks!