Have an idea?

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

Significant differences of attribute levels between groups (HB)


I run an HB analysis giving me the individual-level utilities. To account for possible differences between males (n=182) and females (n=89), I've calculated the mean part-worth utilities and standard deviations of each of the groups (m/f).
I want to ask you whether it is the correct way to "simply" do a t-test by averaging the values for each attribute level (averaging mean part-worths males and females and standard deviation), calculating the other important values (i.e. standard error) and checking whether the final results are +/- 1.96 on the 95% confidence interval.

I would be great to hear from you soon :-)

Best regards!
asked Jul 22 by briniminii (490 points)
retagged Jul 22 by Walter Williams

1 Answer

0 votes
Yes, you can do the test as you've described.  I would be careful to use not the raw utilities but the zero-centered diffs (ZCD), which also come out in the summary report if you're running this in the Lighthouse Studio software.   ZCD is the software's way of ameliorating respondent-level scale effects and making utilities more comparable across respondents or groups of respondents.
answered Jul 22 by Keith Chrzan Gold Sawtooth Software, Inc. (48,525 points)
Thank you!
I've done so, but I think I rather have to take the difference between the average utilities of the two groups, and then I divide the difference of the mean part-worts by the standard error (average standard deviation between the groups/ square root of sample size (n-1)), instead of the average between the mean part-worth utilities, don't I?

Also, for my further analysis can I / should I proceed with this kind of segmentation (f/m) for HB, i.e. testing RLH, Percent Certainty, hold out choice sample and co. for each of the groups, or is it acceptable / correct to test the model accuracy and quality without focusing on each group?
I don't understand your first question - you seem to have the mechanics of the t-test down and you can do that, yes.   A formal statistical test, the Swait-Louviere test, exists for this kind of analysis using aggregate logit (and correcting for differences in scale parameters between the groups) but if you want to use the respondent-level parameters, the way you're describing makes sense to me.  

I don't usually see people trying to assess model accuracy for subgroups, but nothing stops you from doing that if you want to do so.