Have an idea?

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

Evoked (Consideration) set in constant sum allocation CBC

I'm doing a pharma study with a large number of SKUs. Each respondent allocates 100% of future prescriptions among all alternatives. I plan to analyze the data using CBC HB software.

There are two attributes - product and price.

Due to the study specifics, not all of the SKUs (products)  are going to be shown to each respondent (not only because they are unattractive for some of them, but also because they are inaccessible). Therefore, something like an evoked set (consideration set) of SKUs is going to be formed for each respondent.

The questions is, what do I need to write down into the .chs file for the SKUs not present in a respondent's consideration set?

1) First, I could include all of SKUs into all respondent records in the .chs file and artifically insert zeroes as points allocated to SKUs not shown to a respondent. I suppose this is going to decrease utilities of SKUs rarely accessible (which seems to be good idea for predicting market share, not the share of preference). Furthermore, I would need to enter some price levels for such SKUs as well, which would decrease the utilities of these price levels either.

2) Secondly, I could simply exclude irrelevant SKUs from respondent's records, which would result into different number of concepts per tasks for each respondent. This approach seems more logical to me, but I'm not sure if it's correct in terms of the way CBC HB treats .chs data?
I've noticed that CBC HB utility estimation has no problem with such .chs files, but I still have doubts.
asked Sep 7, 2015 by magoli (150 points)

1 Answer

0 votes
Best answer
Very interesting questions, many of which were addressed at the 2015 Sawtooth Software conference in a paper by Belyakov that is not yet published because we're still in production on the Sawtooth Software Proceedings.

My colleague Kenneth Fairchild and I have been conducting some extensive simulations  with robotic respondents on the issues you mention, though we are dealing in discrete choices rather than allocations.  (We plan to present our research at the upcoming "Turbo Choice Modeling Seminar" in Florida in February.)  I believe that what Belyakov and we have found should be applicable to your situation.

I'd recommend the following:

For levels of an attribute that are dropped (in an evoked set design) due to the level being inferior, you can indeed pretend that the respondent also "saw" these concepts available within each choice set (but always get them an allocation of zero).  Regarding prices for these "phantom" concepts that were not actually seen and therefore cannot be chosen by the respondent, you should experimentally design the prices so that they are varying.  If you specify a price of zero for these phantom concepts that are represented in the design matrix for each choice task, you will bias your price utilities.

Regarding the levels of an attribute that were absolutely not accessible to a respondent (and should have 0 probability of being chosen), you should ignore the issue during HB estimation, but post hoc you should manipulate your file of individual-level part-worth utilities to set such inaccessible levels to have utilities of something strongly negative such as -99.  Thus, a product composed of that level for a respondent for whom it is unaccessible will end up with a share of preference of essentially zero, no matter what its price is.  

CBC/HB software can handle data sets where the number of concepts and number of tasks differ across respondents.
answered Sep 8, 2015 by Bryan Orme Platinum Sawtooth Software, Inc. (198,315 points)
selected Sep 29, 2015 by magoli