Have an idea?

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

Moderations Using HB Utilities/Importance Scores

Dear All,

I have a question on the “right/best way” to do follow-on statistical analysis with HB utilities (i.e. the resulting relative importance scores for attributes assessed). I used CBC to assess critical features of a new product and already used Lighthouse’s HB functionality to compute individual-level utilities per attribute and the overall importance per attribute. I would like to use the individual-level importance scores from HB regression for a more in-depth analysis of moderating effects (e.g., differences across genders, different customer types, etc.). However, I am not sure if I just simply do this and, thus, have the following questions:

1)    Can I generally use the relative importance scores (resulting from HB regression) to compute moderations (e.g., is there a statistically significant difference in the importance of attribute 1 depending on the gender of the respondent)?
2)    If yes, would I need to transform the relative importance scores in any way before conducting follow-on statistical analysis such as moderations?
3)    If yes, which statistical method would be the most methodologically sound method to assess whether different variables moderate/affect the relative importance of certain attributes—i.e. would I do this best via regression analysis (and if yes, which type of regression, e.g., logit models) or would conduct tests like one-way ANOVA, Welch test or Kruskal-Wallis H test?

Many thanks for your support in advance and happy Sunday,
asked Jan 23 by DSkambraks (180 points)

1 Answer

0 votes
1) Yes
2) No transformation needed - it's already been done for you
3) check out this article, which has the relevant advice:  https://sawtoothsoftware.com/resources/technical-papers/statistical-testing
The chapter shows Bayesian and frequentist test of hypotheses.  In addition, people looking to publish results of the stat tests you're talking about will typically enter them into a logit model along with the attributes and levels, either as moderating (interacting) variables or directly as main effects predictors (if you've used a labeled conjoint experiment), and they'll usually do so in a way that accounts for account for the logit scale parameter (see the Appendix to the chapter)
answered Jan 23 by Keith Chrzan Platinum Sawtooth Software, Inc. (110,575 points)