Have an idea?

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

Share of Preference vs. Probability based Rescaling Procedure in MaxDiff


When we run the HB estimation on MaxDiff data, we get three sets of scores:
1) Zero centered raw scores
2 Zero centered interval scores
3) Probability based Rescaled scores

For our studies, we use the Zero centered raw scores to calculate the Share of Preference % using the exp(i) / sum of all exp(ij) formula.

We had two questions about these:
1) What is the difference between Zero centered raw vs. Zero centered interval scores?
2) If we REALLY had to pick one between the Share of Pref. % and Probability based Rescaled scores (which also sum to 100), which one would that be and why?

It seems like both the Share of Pref. % and Probability based rescaled scores are calculated based on the same raw scores, but they do give different outputs a lot of times (both in terms of ranking of attributes as well as the percentage value). We want to know which model is more robust and recommended.

Thank you in advance for your valuable time, please let me know if you have any clarifying questions.

asked Jan 13 by kshitijkumarsingh (300 points)

1 Answer

0 votes
Best answer
Replies to both your questions:

1) The zero-centered interval scores just stretch the zero-centered raw scores to that the items have a range of 100 for each respondent.  Some folks think these might be better to use as inputs to segmentation, say, but I think it would be better still to use latent class MNL to create segments directly from the choices, not from the separately-estimated utilities.

2) The Probability-based rescaled scores try to take into account that items in your survey are scaled to a logit model that had (whatever you used, usually 4 or 5 items in a set) whereas the share of preference logit calculation you're doing assumes a (counterfactual) comparison of each item with all the others.  Some folks at Sawtooth Software like the Probability-based rescaled scores better than the straight (counterfactual) logit scores.  You're right that the (summary/mean) conclusions can change based on which (if either) rescaling you use but if there's a definite right or wrong I'm not aware of what it is.  I know smart people who use either (and in fact some who use neither).
answered Jan 13 by Keith Chrzan Platinum Sawtooth Software, Inc. (114,000 points)
selected Jan 14 by kshitijkumarsingh
Thank you Keith.