Have an idea?

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

Minimum tasks/ respondent in CVA?


1. N=325.
2. Assume a grid of 6 attributes, 3 of which have 4 and 3 of which have 3 levels.
3. Using CVA, I am curious as to the minimum number of tasks one would recommend for HB per respondent.
4. I am potentially considering a "1-7" likelihood to use the product on the CVA screen.

I have read the manual on p.806 and also searched the forum. I understand that HB can be done on "large samples" with a sparse design; I understand that the number of tasks/ respondent is (generally agreed) to be very robust (in that it seems to be overkill) if one uses what lighthouse would like you to do.

In the manual and on some sources, I see some statements that "a minimum of 6 tasks" could work for HB if sample is large enough. However, I have yet to figure out what is "large enough." The example in the manual is 500, 1000 and 1500 respondents, for example.

I believe the answer to this is probably somewhat murky in that I believe that it is sort of chicken and the egg: if the attributes and levels tend to lend themselves to a clear winner and obvious trade-offs, the analysis could stabilize with fewer respondents.

However, I would appreciate if anyone has any insight as to going into this without any a priori knowledge of what may come out of the CVA-- what would be the minimum recommended number? Is there a way to understand the benefit of 6 tasks vs. 7 vs. 8, if one chooses a sparse design?

Thank you.
asked Aug 24, 2017 by anonymous

1 Answer

0 votes
Prior to seeing the data, it's really hard to know the answers to these questions (prior to knowing the size of the effects and the variance across respondents).

A lot depends on how large your sample size is and whether you want to draw strong conclusions about segments of your data, or for the sample overall.

Your planned sample size is n=325, which is not considered large.

If using HB estimation and CVA, I generally would want to have 1.5x as many conjoint questions as parameters to estimate (assuming I wanted reasonably high resolution at the individual level).

You have 16 parameters to estimate, so I'd recommend 24 questions.
answered Aug 24, 2017 by Bryan Orme Platinum Sawtooth Software, Inc. (142,715 points)
Thank you Bryan: incredibly helpful.

Let me provide a bit more context (if I may take advantage of a very helpful person): let's say conclusions of sample overall, am using HB, and I would not be worrying about individual level conclusions but rather across the sample.

Does that change your reply?

Thank you for your help-- fantastically helpful.
Thanks for more context.  If I was worried about the length of my survey and really felt I had to trim the length of the CVA section.  Assuming I didn't need to look at segment differences (only the overall sample).  Assuming the n=325 and assuming HB estimation, I could see reducing the number of conjoint questions to 1x the number of parameters to estimate, or 16.