Have an idea?

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

Dual Response Chip Allocation

We fielded a CBC study that included dual-response none questions on a 5-pt scale.  I’m interested in trying to estimate the “allocation none” method here, using the response to the none question as chip allocation.  4 chips for product if chose scale point 5, 3 chips for product, 1 chip for none if chose scale point 4, etc.  Full disclosure I’ve never ran a dual-response study, nor a chip allocation study!

So my question is about task weight in CBC/HB and what value is appropriate.  Since this is dual response, only half the tasks will be allocation, the others discrete.  But maybe that’s fine as the manual mentions the data are normalized anyway during estimation.  So if I’m doing a 4-chip allocation, do I take the manual’s advice and make the weight 2, half way between full allocation and full discrete?
asked Mar 20, 2017 by JKincaid Bronze (1,035 points)

1 Answer

0 votes
This is a good question.  I'm not sure anyone knows the right answer but I can tell you how I would think about it.  First, for the allocations, it depends a bit how respondents are using your five point scale - if respondents tend to be all over the place, giving lots of 1s, 2s, 3s, 4s, and 5s, then they really are using their allocations as independent choices and that might incline me to think that the task weight should be a bit closer to 5 than to 1 (at least as far as considering the allocation questions goes).  But if respondents do what is much more common, and provide strongly skewed answers to the rating scale, then some answers (definitely will buy, probably will buy) are appearing a lot more than others and and your respondents may not be using the rating scale as much more than a dichotomy - which would incline me to weight the tasks much less.  Bottom line, I think your compromise of a task weight of 2 is a good idea, but I might also run it with a weight of 1 and a weight of 3 just to make myself comfortable that my results aren't too sensitive to my chose of one task weight over another.
answered Mar 21, 2017 by Keith Chrzan Platinum Sawtooth Software, Inc. (102,700 points)
Nice response, Keith.  

Just recognize that increasing the task weight will cause the utilities to fit the individual choices better (put more emphasis on the lower level) and less weight on the upper-level model in HB modeling.  That's because greater task weight makes it look like the respondent has provided more data.  

Note that increasing the fit at the individual level is not always a good thing!  It can lead to overfitting.  You want to find a good mix between the influence of the upper and lower-level models in HB.