Have an idea?

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

Alt-specific design in CBC/HB

Hello, I'm not yet Sawtooth customer but before I become him I would like to clarify one point.
According to CBC/HB user manual the system supports alt-specific design (Car/Bus example, p.76). However, I can't find how I should set that some attributes are alternatives, not just regular, experiment-wide attributes. Respectively, there is no way how to tie alt-specific attributes to their alternatives (except zeros in responses).
As result, the output seems to be just betas (assuming flat design) , while for alt-specific design the output for alternatives should be constants to represent unobservable random effects.
Could you comment this? Is CBC/HB system so smart that able to count zeros in responses to  separate alternatives from attributes and correctly ties them to each other?
asked Feb 2, 2016 by Nikolay (145 points)

1 Answer

+1 vote
Thanks for the question.  Often other software packages for discrete choice modeling capture the effect of the alternative "label" as an Alternative-Specific Constant (ASC).  With our CBC/HB software, one creates another factor in the experiment to account for the labels of the alternatives.  Also, we use effects-coding rather than dummy-coding.  These two things are a bit different from what most econ or stats people are used to, but they result in equally good models and the same model fit as dummy-coded designs and those that capture labels' effects as ASCs.  

When a factor (attribute) doesn't apply to an alternative in the design, we zero it out in the design matrix for that alternative.  Effects coding zero-centers the effects within each factor to achieve identification (rather than fixing the reference level to zero within each factor).  Either way works fine and results in the same model fit...it's just that many people find the effects-coded parameters are easier to interpret than the dummy-coded ones.

So, if you follow our documentation in the CBC/HB software for alternative-specific designs and just explicitly code your ASCs as levels of an additional factor, it all works out properly.
answered Feb 3, 2016 by Bryan Orme Platinum Sawtooth Software, Inc. (154,305 points)
Excellent, crystal clear answer. Thank you, Bryan.