Have an idea?

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

Advice on designing for a second 'message' from the same set

An attribute ('Claim') has levels that could appear either once or twice in each choice.

We think that the impact of a second claim will be less than the sum of the individual appearances (however, the order of preference should be the same - identical claims will be excluded).

Is there a straightforward way to set up the design for this?

The approach I'm planning initially is to copy 'claim' and then merge the two variables back down into 1 variable in the cho file. Will the software be able to CBCHB software cope with multiple items being available in one attribute?

Update - I'd confused the cho file format with the design matrix.

So given that ... Is there a way to directly specify the design matrix such that I can indicate:

Two levels are (sometimes) present at the same time (and potentially separately that there is an interaction between the two levels that makes their presence less than or more than additive)?

I guess the model I want to specify is U = Sum of level U's x AModifier for two being present
but I can settle for the traditional U = Claim 1 level (if present) + Claim 2 level (if present) if there is no trick to enable this.
asked Mar 17, 2015 by Andrew Reynolds Bronze (1,140 points)
edited Mar 18, 2015 by Andrew Reynolds
Mistaken edit

1 Answer

+1 vote
Not 100% sure if I understand, but it sounds like you can double up on the "claim" attribute twice,  use prohibitions to prevent the same claim from showing twice.  Normally you'd use an interaction effect to capture any extra boost or loss, but the presence of prohibitions will prevent this in the software.  You would need to probably recode the two attributes down into a single attribute with all relevant combinations in order to inherently capture the interaction.  This could be done by modifying the .cho file, or by saving a copy of your survey, modifying the design in Excel and inside SSI Web, and importing the design via a CSV import.  Your answers will still line up with what was displayed to respondents.

Not going directly for the "one attribute with all relevant combinations" approach gives you the option of sticking with a main-effects only model if desired, though then you have the problem that you get two utility estimates for the same thing.  It kind of feels like you're capturing an order effect this way, but I'm not sure how confident I feel about that.  You could potentially average the two scores together if you wanted to figure out utilities and importance scores like you're used to, but there might be some consideration I'm not immediately thinking about.
answered Mar 19, 2015 by Brian McEwan Gold Sawtooth Software, Inc. (41,270 points)
Thanks Brian. You understood perfectly and I'd come to the same conclusion. Unfortunately the number of levels is too high so I'll have to stick to main effects. Claim 1 is shown 2x the frequency of Claim 2 so I 'll have more faith in that. I'll see how Claim 2 outputs before I decide how much use I can make of it. I'm hoping it will be proportional to claim 1 and just differ by scale. If not averaging the two may be the way to go.

In previous studies I've used a holdout to calibrate for a 'repeated' attribute but this has its own issues. I can't think of a clean way to deal with this situation when there are a large number of levels.