Have an idea?

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

Creating an experimental design for CVA with interactions


I have a question that is not (yet) related to a specific Sawtooth product, but rather of a general kind:

I want to do a CVA with 10 factors, each having two levels (i.e. 210), and need to estimate not only main effects, but also two- and three-way interactions. As the full factorial design is rather big (1024 profiles), I considered a fractional factorial design to be a good idea. So far, so good (I think).

As a next step, I considered either using an orthogonal array of strength 7 / resolution 8 or create a D-efficient design that fits my requirements. However, I have a problem with both of these options:

A)    I found two orthogonal arrays of strength 7 that would be applicable, but they need 128 or 1024 runs, respectively. As this number obviously needs to be “reduced” I was wondering: Would I be able to create a certain number of blocks (out of one of these two arrays) to reduce the number of runs for the individual and then be able to estimate main effects on the individual and the interaction effects on an aggregate level? And if this is the case, how would I do this?

B)    As I did not find an answer to the previous question, I wanted to check if I can create a D-efficient (not necessarily optimal) design of resolution 8 with a manageable number of runs for the individual - Warren Kuhfeld suggests using SAS for this task. However, I don’t have access to this software right now. So the questions here are: Do you know if there is any other software with which I could equally create and test the described design? If yes: Which one (SPSS and Stata do not seem to work; R seems to be an option…)? If no: Do you think SAS can really help me with my problem? (In order to better be able to assess if buying this software would be a sensible investment for me…)

In any case, I’m very thankful for any kind of suggestions or recommendations that you might have! Thank you a lot in advance!
asked Apr 19, 2016 by anonymous

1 Answer

+1 vote
Our CBC designer often makes better interaction designs than does CVA.  Just specify that you want one or two concepts per task (depending on whether you want to run a one- or two-concept CVA).  If you have more than the 10  versions that CVA you can still analyze via HB/Reg.  SAS is also a good option if you can spend the money - it is very flexible.  Very affordable is NGENE, though I've found that sometimes it fails to find a simple design I know to exist, and other times (e.g. some Bayesian designs) where it SHOULD be able to make better designs than CBC, but it fails to do so.   Still, it's a lot less costly than SAS and may be worth trying first.
answered Apr 19, 2016 by Keith Chrzan Platinum Sawtooth Software, Inc. (90,475 points)