Have an idea?

Visit Sawtooth Software Feedback to share your ideas on how we can improve our products.

How to handle mutually exclusive attributes in CBC?


I need to come up with a CBC design for a product that has two mutually exclusive attributes (both with 5 levels): if a level from attribute A is shown then attribute B must be blank and vice versa. As I see it, there are 3 ways to do this:
1. Merge them into one larger attribute with 10 levels - this has the drawback that it needs additional sample in order to obtain an efficient design
2. Use a hidden attribute with two levels (A / B) and the use conditional display to show the necessary text. This would work great as far as sample is concerned, but will we run into trouble during market simulation?
3. Partial profile exercise - same drawback as point 1, larger sample needed.

I'm leaning towards point 2 at this point but my main concern is how will this influence final utilities and the simulator.
Also, I would think that the efficiency report I receive from "Test Design" is misleading: whichever way I go there are 10 levels that need to be shown, so in theory all my alternatives preform equally for a similar sample?

What would you advise? Would I get better estimated utilities in points 1/3 compared to point 2? Is any of my 3 alternatives better in regard to sample efficiency.

Thank you,
asked Jan 13, 2014 by Michal Dudek Bronze (525 points)
You can also try attribute specific design made with the Advanced Design Module. It will be essentially the same as your option 2, but you will get a properly coded "cho" file for an analysis and benefit form proper test of this design is SSI program.

1 Answer

0 votes

I believe your options 1 and 2 could be equivalent once you code the data.  I think in either case you're tricking the designer to get the result you want, which is two five-level variables which never appear together and which will be coded with 4 effects-coded variables each.  It is this reality you will need to test and I think to do so you may need to run the model in CBC/HB with artificial data (whereupon I suspect you will find the two to be equivalent).  

So the design you have generate in SSI Web's CBC designer needs to be recoded to capture your model and analyzing the recoded design will be a bit of (probably unnecessary) work for you.

Your third design alternative, partial profile, probably is not a solution for your problem.   However, I do want to point out that people misunderstand the value of partial profile designs - such designs do not (and cannot) compete with full profile designs in terms of statistical efficiency.  Their big benefit is in simplifying designs with many attributes and reducing respondent error, thus increasing respondent efficiency and improving total efficiency (but not, again, statistical efficiency).
answered Jan 13, 2014 by Keith Chrzan Platinum Sawtooth Software, Inc. (69,125 points)