Mixlogit: calculate probability of choice for UNLABELLED alternatives including case-specific regressors

[Julia Schreiner]

Dear Statalisters,

I am analyzing data from a DCE to obtain wta of animal welfare programs. Each farmer was confronted with 4 choice sets comprising 2 programs and 1 opt out. Each program has 5 attributes (3/4 levels) varying over alternatives. Data is in long format as required with levels coded as dummies. Since farmers have heterogeneous preferences I decided to use mixlogit to evaluate the the probability of acceptance depending on alternative-specific AND individual-specific variables.
First I tried only the alternative-specific variables and obtained significant standard deviations for all attributes. I included the attributes as random variable except the price.

mixlogit choice p, rand(attr1 attr2 attr3) id(Index) group (case) nrep(500)

Afterwards I wanted to estimate a model with individual-specific variables. Referring to Hole (2007) and Cameron & Trivedi (2009) this is possible by creating interaction terms between the attributes and characteristics of the decision maker or by creating interactions between alternatives and characteristics. Since I have unlabelled alternatives (1 and 2) I want to estimate the probability to choose alternative 1 OR 2 in relation to the opt out (3) depending on the farmers' characteristic.

The first try was to generate a dummy if alternative 1 or 2 was chosen and created an interaction term between the dummy and the characteristic, BUT the model achieved no convergence.

I recently worked with asclogit and used constraints to set alternative 1 and 2 equal. How is it possible to do this when using the mixlogit command?

I am really looking forward to helpful comments! Thank you so much in advance,

Julia

[Arne Risa Hole]

Julia

To include individual-specific variables they should be interacted with the alternative-specific constants, not with the choice variable. For example, if you have a dummy for being female in your dataset, and you want to include that in the model subject to the constraint that it has the same impact on choosing alternative 1 and alternative 2 you can run:

gen female_int = female * (Alternative==1 | Alternative==2) if female <. & Alternative <.
mixlogit choice p female_int, rand(attr1 attr2 attr3) id(Index) group(case) nrep(500)

Note that you can use the same approach with asclogit instead of using the constraints option - see http://www.statalist.org/forums/foru...using-asclogit

Arne

[Anat Tchetchik]

Dear Arne,
I'm trying to include interaction var. in -mixlogit- using the operator # (for example i.var1#c.var2). Yet, Stata returns:

factor variables and time-series operators not allowed

. Should I manually create the interaction term? one last quest. that I I'n sure about is whether interaction vars. (for example between an att. and respondent's demographic) can be included as a random term in the -mixlogit- model under the rand( ) option.
Thanks,
Anat

[Arne Risa Hole]

Anat
Yes, you will need to manually create the interaction variables before including them in the model. Interactions can also be given random coefficients, i.e. included in rand().
Arne

[Julia Schreiner]

Dear Arne and other Statalisters,
I have another question concerning the estimation of a RPL model with mixlogit package. This time I have three alternatives (no labels) and no opt-out (forced choice). If I would create a constant for all three alternatives, I would not have within group-variability. Two options came to my mind: the first one would be to look which alternative contains in total most base levels for each attribute and chose it for the reference category OR to generate interactions only with attribute levels. Are there any other ways to estimate a forced choice setting? I'm thankful for any help!
Best wishes, Julia