Have an idea?

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

interpretation HB average utilities

Dear Sawtooth Software Team,

I have a few questions regarding the interpretation of CBC HB part-worth utilities for my master thesis.

1)  a negative average utility value DOES NOT say that the level is providing negative utility. Correct?
For example: An attribute has 4 levels. Utilities of level A and B have a negative sign. But that does not mean that A and B is disliked, right? We can not interpret the sign.

2) Can I report at which rate utility is increasing between levels? For example: Change from level A to B is 70. From B to C is 30. From C to D is 12. Can I state: Further increasing attribute X is still adding utility but the at a lower rate?
Can I interpret the utility differences between levels?
For example, one level is regarded more negative than another since there is a utility difference of xy.
Suppose there is level A, B and C. Utility increase between A and B is very low / their utilities are similar. But utility increase from B to C is high. Can I then say that compared to level C, A & B are rather disliked?  Also, can I say: utility increase is highest when increasing B to C?

4) Also, when I compare groups of people (e.g. demographics): Is it correct to say something like: Men drive more benefit out of level xy (∆U=7; p<0.05)? (For these comparisons I use the zero centered individual utilities and compare means via t-test)
related to an answer for: estimated utilities
asked Sep 16, 2020 by anonymous
edited Sep 16, 2020

1 Answer

0 votes
Any update here?
answered Feb 20 by danny Bronze (1,260 points)
Indeed, because utilities are zero-centered and therefore, utilities are interpreted in a relative sense (negative utility means relative worse than other levels within the same attribute).  Of course, respondents may think all levels are very good...but some levels have to be relatively lower than the others.

The raw HB utilities are on the logit scale.  That means that those utilities exponentiated are proportional to choice.  Or, stated another way, the logit-scaled utilities are natural logs of probability-scaled scores.  Yes, you can say that relative to C, A & B are less liked.  And, on the logit scale, increase is less between A and B compared to either of those to C.

When comparing groups of people on a normalized scale (such as zero-centered utilities), you can say that one group of respondents prefer a level more relative to the other levels within the same attribute as another group.