Aggregating attribute levels into single attribute in DCE analysis

[kcchia95]

Hello,

I am estimating a discrete choice experiment (DCE) using a Mixed Logit model (preference space) and would appreciate your advice on a modelling decision.
One of the attributes, award, has 5 levels: 4 different types of awards and the absence of an award. The original experimental design was generated with Ngene, the code is as follows, and the b5.dummy represents the award attribute:

Design ? Bayesian D-efficient design for GWC survey (INTERACTION)
;alts = choice1*, choice2*, choice3*, choice4*, choice5
;rows = 32
;eff = (mnl,d,mean)
;bdraws = halton(100)
;block = 4
;model:
U(choice1) = b01[0]
+ b1.dummy[(n,0.1,0)|(n,0.1,0)]*A[1,2,0]
+ iae1[(n,0.1,0)]*A.dummy[1]*E.dummy[1]
+ iae2[(n,0.1,0)]*A.dummy[2]*E.dummy[1]
+ iae3[(n,0.1,0)]*A.dummy[1]*E.dummy[2]
+ iae4[(n,0.1,0)]*A.dummy[2]*E.dummy[2]
+ iae5[(n,0.1,0)]*A.dummy[1]*E.dummy[3]
+ iae6[(n,0.1,0)]*A.dummy[2]*E.dummy[3]
+ iae7[(n,0.1,0)]*A.dummy[1]*E.dummy[4]
+ iae8[(n,0.1,0)]*A.dummy[2]*E.dummy[4]
...
+ b5.dummy[(n,0.1,0)|(n,0.1,0)|(n,0.1,0)|(n,0.1,0)]*E[1,2,3,4,0] ? 5. award
+ b6[(n,-0.1,0)]*F[1,2,3,4]
/ ...

I have collected the data, and I am considering conducting two approaches for the model estimation:
In the first specification, each award type is estimated separately.
In the second specification, I am considering replacing the four award type levels with a single binary variable indicating whether an award is present (1 = award present, 0 = no award), regardless of the specific award type.

My questions are:
1. From a methodological perspective, is it acceptable to aggregate the four award types into a single “presence of award” variable?
2. Are there any statistical or experimental design issues that I should consider, such as the imbalance of the two levels (presence/absence), the "aggregated" attribute, and potential biases?

Thank you very much for your time and advice.

Best regards

[Michiel Bliemer]

It is perfectly fine to aggregate dummy coded categories during model estimation, and this is often done in situations where the coefficients of individual categories are not statistically significant or have ordering issues. So converting 1,2,3,4,0 into (1,2,3,4)->1 versus 0 should not be a problem.

Your experimental design was optimised for estimating parameters for a 5-level categorical attribute, but you can use the data to also estimate parameter for an aggregated 2-level attribute. A design optimised for a binary award attribute would have been more efficient, and would have had more balance in the data across award and non-award. You current data will have more award (1) observations than non-award (0) observations, but this should not bias your results.

There should be no issues in aggregating categories during model estimation.

Michiel