Have an idea?

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

None weight in ACBC

I can read some general advice for the none weight but not very specific for ACBC. Here the none parameter seems to have a completely other foundation than in CBC (threshold estimation)

Has anybody experience what is "allowed" with none weights in terms of size?
asked Aug 15, 2018 by anonymous

1 Answer

0 votes
Indeed, with ACBC the "None" weight that automatically comes out of the standard ACBC setup (BYO, Screeners, Tournament) is a different animal that what comes out of a CBC.  With the standard 3-section setup in ACBC, the None weight is actually the utility for the "not a possibility" alternative in the Screener section.  

Some researchers like to add the optional "Calibration" section to the end of ACBC surveys (BYO, Screeners, Tournament, Calibration) and then ask our software to do the additional step (after estimating HB utilities) of calibrating the None utility so that it fits (in a least-squares sense), say, a 4 on the 5-point purchase intent scale for each respondent.  

Regarding what is "allowed" in terms of size of the parameter, if using the standard None coming from the "not a possibility" option, it is whatever the MNL estimation allows within the HB estimation.  In my experience, if I am using Summed Pricing with the piecewise function for price without monotonicity (sign) constraints on the price attribute, then the None weight can wander around quite a bit across the history of HB estimation.
answered Aug 15, 2018 by Bryan Orme Platinum Sawtooth Software, Inc. (203,065 points)
Thanks  a lot for your answer. I observed that average partworth of "none" can vary quite a bit depending on contraints. But then it is not a big difference when used in simulations. That's fine.

What I am  wondering is what none weight I could apply in SMRT.
In my example I use a static straight forward choice question of 3 products in the questionnaire (allowing "none of these") - based on a combination of attribute-levels. I call that calibration question.

When I build these products in SMRT using the partworths I get results that look fine (the three share of choices relative to each other) but the none share is quite different. When I use none weight as low as 0.005  it gets about OK.

My question was whether this is not recommended.

Maybe there is a logic like "by random" In the screening section I get 50% none - in a Simulation this looks different as there are more products...
Using just one CBC-looking question with a None alternative to calibrate ACBC's None utility within SMRT to predict that one holdout question is placing a lot of trust on a single choice task observation.  I would think you'd need at least 5 such CBC-looking holdout tasks to make such an adjustment.

It is true that the "None" scaling from the binary comparison of each product in the Screening section to the "Would not work for me" is quite different from the scaling of the None when shown as an alternative in standard CBC-looking sets of product alternatives.