Have an idea?

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

ACBC and share of preference simulation: Additive model of part-worth utilities appropriate if cut-off rules are used?

I am using ACBC in order to build up a share of preference model based on the HB estimated part-worth utilities. I am not using the SMRT, the part-worth utilities are used in an excel model with Logit or Maximum Utility rules in order to calculate share of preferences.
Given that cut-off rules (asking must-have and unacceptable questions) are used in the screener section of ACBC I was wondering whether it is appropriate to use an additive model in order to calculate the final product alternative utilities.
Literature says that an additive model requires compensatory decision rules which are not the case if must-have questions are in place.
Does the HB procedure account for cut-off rules or do we have to adjust the estimated part worth utilities according to the must have and unacceptable results?

Regards
Fabian
asked Nov 15, 2011 by anonymous
retagged Sep 13, 2012 by Walter Williams

1 Answer

+1 vote
 
Best answer
An interesting thing about the additive rule is that it can accommodate non-compensatory cut-off rules quite easily if the utilities are made extreme enough.  As a simple example, if all the utilities are in the range of -3 to +3, but the respondent hates brand x... if brand x is given a utility of -100, then nothing can be done to add up utilities across the other attributes to compensate for the fact that the presence of brand x kills any likelihood that this respondent will pick a product with brand x.

So, you can see that if some utilities (representing "must-have" or "unacceptable" levels) are scaled extreme enough, the additive rule can correctly reflect non-compensatory decision-making.

While ACBC indeed will compute quite extreme good or bad utilities on levels that are must-haves or must-avoids (using HB estimation), we are reluctant to make the utilities wholly non-compensatory in nature.  Too much experience and evidence over the years with human respondents suggests that many people will indicate that a level is a must-avoid or a must-have, when in reality their decision rules are not quite so fixed.  So, the HB estimation process does some smoothing to the population means, even for extreme utilities, making the resulting extreme levels less extreme than they would have been if we take those cutoffs as absolutes, leaving open the possibility of trading into these alternatives if another really good features are present within the product concept.

So, in short, you may indeed use the additive logit rule (or first choice rules) in simulations based on ACBC questionnaires.  It works out well.
answered Nov 16, 2011 by Bryan Orme Platinum Sawtooth Software, Inc. (148,340 points)
Bryan, could you briefly explain how the extreme scaling in utilities according to cut-off rules is done in the estimation?
Furthermore I was wondering if using constraints in ACBC isn’t of any use then, if the extreme scaling of utilities represents cut-off rules. Wouldn’t constraining the utilities especially in ACBC diminish this intended effect?
In ACBC, we generate a bunch of "near-neighbor" product profiles that are similar to the respondent-specified BYO product concept.  Let's imagine for a particular respondent and study there are 24 such near-neighbor cards.

Now, in the Consideration phase of the interview, imagine I as a respondent indicate for the first 12 cards that any profile that is Brand X is not a possibility.  The Unacceptables question probes me, to confirm that Brand X indeed is unacceptable.  Imagine I confirm the rule.  Then, ACBC looks ahead to the 12 remaining cards I haven't yet answered.  If 4 of those are cards with Brand X, then ACBC automatically marks those 4 concepts as "not a possibility" for me, as if I had seen those concepts and marked them as not a possibility in the Consideration phase of the interview.  Then, it generates four replacement cards that are not Brand X to take their place.

Then, utility estimation occurs naturally, per an effects-coded design matrix and the logit rule.  There isn't anything special and additional that we do to scale unacceptable levels.  By virtue of the fact that every time Brand X appears on a card, that card is viewed as "not a possibility" (not chosen in that binary logit aspect of the data), the part-worth utility for Brand X is driven downward.  For that respondent, it has a likelihood of choice of near zero.  If it weren't for HB, the part-worth utility would try to run off to negative infinity.  But, due to the smoothing and population information within HB, that doesn't occur.  So, it becomes strongly negative, but usually something like -5 to -10, rather than negative infinity.

Constraints shouldn't affect the scaling of the unacceptable levels, unless of course the constraint tried to enforce that the unacceptable level was somehow better than an acceptable one.
...