Have an idea?

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

MaxDiff for refining attributes & Levels for a CBC

Hi,

I'm looking to use a maxdiff to refine a list of attribute levels before designing a partial profile CBC. My plan is to use the maxdiff to identify irrelevant Attributes/levels (i.e. ones that are least likely to be determinants of choice) then look to remove those attributes/levels which are not usefull so that my CBC design is tighter. I have about 70-80 items that I'm starting with. Are there any rules of thumb I should be using with this number of attributes? I've not done a maxdiff with so many attributes before so I'm not ensure if there is something specific to look out for. We have already done some qual work to establish the attribute and levels but I believe these can be refined further using the maxdiff.

Any help on rules of thumb (sample size, design tips etc.) would be greatly appreciated!!

EDIT: I have read in another post that Bryan likes to have each attribute spread across the sample 1000 times if I was to simply use aggregate logit estimation.

If that is the case would I be correct in assuming a design with 14 Choice sets of 6 attributes per set over 1000 respondents if I have 80 attributes be a suitable solution?
asked Apr 27, 2017 by Jasha Bowe Bronze (1,680 points)
edited Apr 27, 2017 by Jasha Bowe

1 Answer

+2 votes
Jasha,

Conducting MaxDiff in a preliminary research project is a great way to think about narrowing the list to focus on the more important aspects in a follow-up CBC.

Our software recommends that each item appear 3x per person in a MaxDiff questionnaire.  But, this assumes researchers want and need stable individual-level estimates via HB.  If you dig further into our documentation on MaxDiff you will find that we also recommend sparse MaxDiff for aggregate analysis, when the researcher doesn't necessarily need stable individual-level results (e.g., each item might appear just once per respondent).  Such can be analyzed using aggregate logit or even low-dimensional Latent Class solutions.

I general, when thinking about aggregate (pooled) analysis, I recommend to people that each item should appear at least 500 times and preferably 1000 times across all respondents and MaxDiff sets.
answered Apr 27, 2017 by Bryan Orme Platinum Sawtooth Software, Inc. (139,965 points)
Allow Designs to Lack Connectivity
...