Have an idea?

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

Using ACBC: Does HB estimation rely on additive utility model despite non-compensatory asnwering patterns (Screener)

Dear Sawtooth-team,

I have a question regarding the estimation of partworths or generally utiles for an ACBC survey using HB. Going through the various technical papers provided on this topic (ACBC techpap 2014,CBC/HB Techpap v5, 2007 Sawtooth proceedings, Otter p.111ff) I am unable to find a formula for the underlying utility / preference model for an arbitrary stimulus within the choice sets. Using classic CBC, the utility model is additive (compensatory) . Since ACBC includes must-haves and unacceptables, I would expect to see a non-compensatory/ multiplicative utility model , i.e. a single partworth becomes zero by unacceptable rule, the whole stimulus' utility becomes zero. Following the below statements from the mentioned papers, I assume, all three sections are coded as simple choices similar to CBC (Otters scaling optimization between sections aside). Must-haves and unacceptables are regarded as "rejected concepts" where utility is below the "None" threshold and the chosen concepts' utility model is the sum of the relevant utility models (partworth/vector) per attribute, i.e. compensatory estimation.

"The information in the core three sections of the ACBC questionnaires can be coded as a sequence of choice tasks and may be estimated using maximum likelihood estimation under the MNL model" (ACBC techpap 2014)

"add up the part worths (elements of ßi ) for the attribute levels describing the kth alternative (more generally, multiply the part worths by a vector of descriptors of that alternative) to get the ith individual’s utility for the kth alternative" (CBC/HB techpap v5)

"The discussion so far has ignored the rejection of alternatives by rules. Consider an observed response to a rules question (see Section 2) causing multiple alternatives left in the pool to be jointly marked as rejected." (2007 Sawtooth proceedings, Otter p.111ff)

"In the case of ACBC, any information extracted from all three parts of the interview pooled across respondents implies the transfer of information from one person’s SCREENER data to another person’s CBC data. This is because what is unacceptable or a must-have is very likely to differ across respondents. Thus, a model that suitably connects all three parts of the interview is essential for pooling across respondents to make sense." (2007 Sawtooth proceedings, Otter p.111ff)


Many Thanks
asked Jun 8, 2017 by Luke (280 points)
retagged Jun 8, 2017 by Walter Williams

1 Answer

0 votes
Luke,

Thanks for researching this and asking about it!  Indeed, we use the additive, compensatory rule model to estimate across all sections in an ACBC survey: BYO, Screeners, and Choice Tournament.

For each section, we create choice sets to encode the information about people's preferences.  We stack the data to submit to HB-MNL estimation.

Even though we ask respondents about must-haves and unacceptables in the Screener portion of the ACBC survey, we use such data to construct (typically) a few additional Screener choice tasks for each respondent where concepts containing unacceptable levels are always rejected.  These choice tasks are appended to the other explicit choice tasks that the respondent is seeing and all submitted to additive MNL.  For example, if a respondent tells us that the color "Red" is unacceptable, we look across the conjoint cards (concepts) that the respondent has not yet seen and pre-answer any of these cards that contain the level "Red" as "not a possibility."  Then, any of these cards that have been pre-answered due to a non-compensatory rule are replaced with new concepts that don't contain the level "Red".  That way, the respondent still explicitly views and answers the same number of concepts as any other respondent.  But, we now have both explicitly and implicitly answered choice tasks in the Screener section.

I can send you a PowerPoint presentation that gives more detail about how each section is encoded.  And, you can follow up further with me via email.

bryan@sawtoothsoftware.com
answered Jun 8, 2017 by Bryan Orme Platinum Sawtooth Software, Inc. (142,515 points)
...