Have an idea?

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

covariate estimation (acbc)

First of all I sincerely would like to thank you for your help and also for your patience for my repetitive question and for "my covariates".  I really am trying to understand these processes that you have described. But I am still not able to completely grasp it.

When I apply the Bayesian approach, I should be examining the alpha file, however how should I make the comparison between groups, as I need to document the interpretation of these. Should I just look through the output after convergence point, as you have described before, as  <0 or >0 a and come up with the confidence levels?

And for the frequentist approach, do i need to take the range between attribute levels on zero-centered part-worth utilities ? to be able to test attribute importances? also by part-worth parameters do you mean the average utilities on summary table?

My apologies again for this repetitive and very simple questions but I am just having difficulties to get this.

Thank you.
related to an answer for: Swait -Louviere Test on ACBC
asked Dec 12, 2014 by Mir (400 points)

1 Answer

0 votes
Although the Bayesian approach to interpreting how the covariates interact (serve as betas) to affect the individual part-worth utilities is the formally Bayesian correct way to do things, this may prove challenging to explain to your readers and for your purposes.  You may decide it's just easier to go the Frequentist route and to collapse the weight of each attribute into its "importance score" rather than work from the individual utilities of each level in your study.  (If you have a reviewer/advisor, please check with that individual to make sure the Frequentist approach will suffice.)

If you are working from normalized (zero-centered diffs) utilities, then the difference between the best and the worst levels for each attribute is a fine proxy for the "importance score".  If you are using the importance scores that are normalized to sum to 100% across attributes within each respondent, that will give you the same result (just on a different magnitude scale).

Usually when I'm referring to part-worth utilities, I'm referring to the individual-level utilities as output by our HB software to the .csv file...not the average table of utilities presented as a summary report.
answered Dec 12, 2014 by Bryan Orme Platinum Sawtooth Software, Inc. (198,315 points)
covariate estimation (acbc)