How to do a Choice Design with Disallowed Combinations?

[#1]

I am using the DOE Choice Design function to create a design for a discrete choice experiment. I would like to disallow some combinations of attribute levels due to the fact that they don't make logical sense together, and I would like to decrease the cognitive burden on the respondents. (Ex- disallowing the combination of "no jobs lost" and "8 in 10 households struggle financially") 

I am aware the custom DOE option allows for disallowed combinations, but I still want to be able to implement the prior means/variance matrix as permitted in the "choice design" option. As far as I can tell, using these priors isn't possible in the custom DOE and disallowed combinations isn't possible in the choice design tab. Is there a way I can disallow combinations while still being able to implement my prior means/variance matrix?

[#2 · Accepted solution]

Hi @jlynn21 ,  You are correctly aware of the limited functionality available (as of the release of JMP 19.x); unfortunately Choice Design is the only place where you can incorporate these priors, but the functionality for disallowed combinations is not currently supported (to the extent that I researched and tested this).  I highly encourage you to make a bid for this on the JMP Wish List - JMP User Community!  In the meantime I'll continue to search around and if I can provide any additional suggestions I'll be sure to let you know.   Other than simply generating your Candidate Set and using Data Table Filtering (JMP 19 has improved functionality in this direction) or JSL Scripting to filter out the rows for which the disallowed combinations are applicable. 

[#3 · Accepted solution]

Hi @jlynn21,

@PatrickGiuliano already informed you about the current status of the platform and its options.

Choice designs may be seen as a specific subset of Fractional Factorial Design, with only categorical factors, a blocking/grouping of experiments (to create choice sets), a discrete choice response and often some constraints (for example the number of factors/attributes that can change in a subset of experiments/choice set). 
Regarding your situation, I think there might be still two options to consider to help you (that can be combined):

1. Use the Custom Designs platform, and based on the information you already have about your prior study, either use an A-Optimality criterion with A- Optimality Parameter Weights, to place more weight on the effects you want to reduce the variance. You can also specify directly in the Custom design platform the disallowed combinations (through filtering or scripting). Adding a random block can help creating the choice sets with the desired size.
2. If you already have an existing prior study/choice design, use this design and use the Augment Designs platform. This way, the information from your previous design is kept, and the augmentation will help estimate more precisely the effects (with a D- or A-Optimality criterion, and possibility to add A-Optimality Parameter weights).

Hope this workaround can help you,

[#4 · Accepted solution]

Just to give a little bit more context for the approaches proposed herein: 

1. Use the Custom Designs platform, and based on the information you already have about your prior study, either use an A-Optimality criterion with A- Optimality Parameter Weights, to place more weight on the effects you want to reduce the variance. You can also specify directly in the Custom design platform the disallowed combinations (through filtering or scripting). Adding a random block can help creating the choice sets with the desired size.

This could work but only if you are okay with not using your prior information and just assuming that the utility is equal across all profiles (i.e. the betas/utility parameters are all zero). Under the assumption that the utility is equal across all profiles, the information matrix for a discrete choice design is equivalent to the information matrix of a blocked design (they call this the utility-neutral design in the paper attached).

So, per Approach 1, you may use the Random Blocking feature in Custom Design to create choice sets of equal size, along with incorporating their constraints.

2. If you already have an existing prior study/choice design, use this design and use the Augment Designs platform. This way, the information from your previous design is kept, and the augmentation will help estimate more precisely the effects (with a D- or A-Optimality criterion, and possibility to add A-Optimality Parameter weights).

This would also work but given you are ok with not using your prior knowledge on the utility parameters, and so, you are assuming utility is equal across all profiles. In some way you are incorporating that prior knowledge (indirectly) since you would be leveraging your existing design (created on the basis of those utility parameters), but you would be assuming equal utility accross all profiles for the augmented runs. 

Said another way: the incorporation of the prior knowledge would be indirect, since without explicitly providing the prior knowledge of the utility parameters it would be hard to guarantee that the options within the augmented choice sets are utility-balanced. The augmented runs would be selected to complement the existing design (i.e. trying to achieve equal replication of factor combinations, etc.), but there would be no guarantees on utility-balance with respect to the utility parameters.