CNL unlabeled design

[LuisPilli]

Hi,

I have a SP data set from an unlabeled design and I want to specify a CNL model.

The attribute supporting nests is brand, which can be unavailable or appear more than once in any given task. Thus, any nest can also be unavailable.

I do not see any straight way to specify this model from the manual and examples. Is there any previous experience that can orient me in this specification?

Regards,
Luis

[stephanehess]

Hi

can you be a bit more specific, maybe by giving us an overview of your proposed CNL structure? How many alternatives do you have, how many nests will there be, etc

Stephane

[LuisPilli]

Hi Stephane,

Thanks for your prompt reply.

I have 6 brands (A-F) as an attribute in an unlabelled design, i.e., one brand can be absent or it can appear more than once in any task. Choice tasks have 6 alternatives.

The simplest structure that I would like to test would have four nests, which would be AB (low-tier), CD (mid-tier), EF (high-tier), A-F. My objective is to estimate the probabilities of different consideration sets, allowing one of the consideration sets to include every brand. It means that every brand maps into 2 nests (low, mid, or high AND A-F).

Is this explanation better?

cheers
Luis

[dpalma]

Hi Luis,

I am afraid that there is no straightforward way to code your model in Apollo. What you are proposing is that the nesting structure changes from one observation to the next. For example, in the first observation alternatives 1, 2, 3, 4, 5, and 6 might be branded A, B, C, D, E, and F, respectively, so the nesting would be low={1, 2}, mid={3, 4}, and high={5, 6}. But then in the second observations alternatives 1, 2, 3, 4, 5, and 6 might be branded A, A, B, B, F, F, so the nesting would be low={1, 2, 3, 4}, mid={empty}, high={5, 6}. You cannot do that in Apollo, as the nesting is assumed to remain the same across observations.

The way I see it, you have two possible ways to model this.

• The first approach would be using a Mixed logit with error components. So you create three error components, one for each "nest". Then, in the utility of all alternatives you multiply these error components by dummy variables, where only one of the dummy variables takes value 1 at a time, depending on the nest the alternative belong to. So the utilities would look something like the following:

  Code: Select all

  V[["alt1"]] = ... + (brand_1 %in% c("A","B"))*EC_low + (brand_1 %in% c("C","D"))*EC_mid + (brand_1 %in% c("E","F"))*EC_high

  Where EC_low, EC_mid, and EC_high are the error components associated to each "nest", so they would be draws.

• The second approach is more complicated, but would allow you using CNL. It would require you to reshape your data thoroughly. Let's imagine each brand can be repeated up to two times in each choice set. Then you can create a database where each choice has 12 labelled alternatives, but only six are available at a time. The alternatives would be A1, A2, B1, B2, C1, C2, D1, D2, E1, E2, F1, F2. If in your original data observation 1 has alternatives branded A, B, C, D, E, and F, then in your reshaped data observation 1 would have alternatives A1, B1, C1, D1, E1, and F1 available, and all other alternatives not available. If the second observation in your original data has alternatives branded A, A, B, B, F, F, then in your reshaped data you would have alternatives A1, A2, B1, B2, F1, and F2 available, and all others not available. This would allow you to nest alternatives as follows: low={A1, A2, B1, B2}, mid={C1, C2, D1, D2}, high={E1, E2, F1, F2}.

Both of these approaches can be easily expanded to account for your overlapping of nests (i.e. where alternative A could belong to either low or mid).

Best wishes
David

[LuisPilli]

Thanks, David!

This was very useful, indeed.
The solution through the mixed logit is quite interesting. But I think that the CNL would allow more straightforward tests of my ideas.

cheers
Luis