Have an idea?

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

Max diff using constructed list

Is anybody experienced to run a maxdiff with a preselection of items.
I.e. Show 100 items and let the respondent select the 30 that are relevant for him - then ask maxdiff questions for these 30.
Something like an adaptive maxdiff.
Thanks and regards.
asked Jul 15, 2013 by anonymous

1 Answer

+2 votes
You could do this easily by building a constructed list based on whatever logic you like.  If you want a maxdiff with the 100 and then 30 you can use the Script functions MAXDIFFRANKVALUE to get the top 30 from the first MaxDiff Exercise.  If you want checkboxes, just use the function AIC.

Then just assign the constructed list to the MaxDiff Exercise.
answered Jul 16, 2013 by Mike Lodder Gold (23,395 points)
Mike is assuming you're trying to build a list based on anwers to MaxDiff, when he refers to the instruction MXDIFFRANKVALUE, which isn't really what you're after.  You are trying to prescreen items into a MaxDiff.

The trick is that each person has to have the same number of items in the constructed list (say, 30).  There is documentation in the MaxDiff manual for this (help + contents + maxdiff/help + creating a maxdiff experiment + developing a list of item).  And, the second concern is that you need to do something post-utility estimation to somehow convey the information regarding what it meant to have an item missing from a respondent's list.  If the items are chosen at random, then the scores impute population means (via HB) for the missing items.  I suspect this is not what you intend to do.  But, if the items were dropped from the list because they are somehow inferior for the respondent, then you'll want to modify the scores using your own means to somehow convey that the missing items for each respondent have some arbitrarily-chosen lower value.  How exactly that should be done is an open question.  So, this is a tricky area and you should proceed with caution.
The idea was to use on the fly scores that result for each individual and put the non-selected items on a score zero. So it is a purely individual level measure.
My understanding is that HB estimation would not just impute missing values with population means, but rather replace it with the mean from a group of people who is similar to that specific respondents in terms of preferences on overlapping  items. For that reason there should be a reasonable overlap in items shown between respondents so that HB would be able to derive similarity between respondents on the overlapping part of items and derive appropriate means for missing items.