Have an idea?

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

MaxDiff HB results and export data manipulation

Hi there,

I am currently analyzing the results of my data using the analyzer in the sawtooth software as it is hard to deal with the .CHO file exported from the MAxDiff study on any other statistical software. (Yet mixlogit is not an option in the analysis and if I want to try it I have to find a solution for that, suggestions are welcomed too)

My question is that If I am aiming to have a model out of the HB analysis which result of the 3 provided should I use? (probability scale, Zero centered or zero anchored?) and how can I use it to get my model??

Note: My study is a Case 2 Maxdiff and the analysis run takes about 29 hours each time & I am waiting for the results of my first model in about 6 hours from now :), I did not know that in advance. My study had 640+ respondents answering 10 states each, total number of items were 31 (with restrictions that items of the same dimension won't be shown together) and each task has 6 items (since the 31 items are divided into 6 dimensions) each item was shown at least twice per respondent.

Thanks a lot & regards
asked Dec 12, 2019 by AMYN Bronze (2,980 points)

1 Answer

+1 vote
You're right that the software doesn't provide mixed logit using the method of simulated likelihoods, but HB is a program for Bayesian mixed logit.

The probability scale is a particular transform that's most useful for Best-Worst Scaling Case 1 Scaling, and I think you may not find it very useful for Case 2.

For the other two it really depends  how you want to use the utilities.  I usually start with zero-centered utilities and then if I want to give my Best Worst Case 2 model a more conjoint-style coding, I also zero-center the utilities for each attribute.  More often I leave the attributes all on the same scale with the same zero, because one of the main benefits of Case 2 is the ability to make cross-attribute level comparisons, and you lose that ability if you zero-center each attribute individually.
answered Dec 12, 2019 by Keith Chrzan Platinum Sawtooth Software, Inc. (111,275 points)
Thanks again for the advice.
I understand that the Zero-Centred utilities are the once most suitable for the application I want which is generating utilities for each level.
My following issue will be that I want the sum of the best levels of each attribute to be equal to 1 which is the best possible case. Then the other levels will be valued in negative and according to the same scale which converted the best levels into 1, how can this be achieved?

Of course you will want to leave the utilities at their raw logit score for any simulations you might do, but for purposes of display, this seems like a simple linear transformation.  I would probably transform from zero-centered to this maximum of 1 .

I have generated the report for the HB Zero-Centred Row utilities. I was expecting to find one of the items to have the value of 0 being the reference item, which I did not see, as all the items had specific values for each respondent. What am I missing here, or what did not I understand correctly?
How can I, later on, use these utilities to generate a single utility for each item? Is that done by the online maxdiff analyzer or by just merely calculating the mean of the raw utilities?

When the utilities come out of the analysis in their rawest form, the last item has a utility of zero.  But it's not usually the item exactly in the middle of the utility scale, so when we zero-center all those most-raw utilities, the zero moves to the exact middle, which, as you note, is rarely exactly atop one of the attributes.  once we zero-center, no particular item needs to have a zero utility.
thanks for the explanation.
Constrains & Covariates in MaxDiff HB analysis