Have an idea?

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

Swait -Louviere Test on ACBC

I really need your help once more!

I want to compare the differences in preferences between respondents. I have included 4 item likert scale questions into my survey to see whether the individualistic or collectivistic character of the respondent would influence her preferences  for attribute choice. I have previously used these 4-item Likert scale questions as covariates in my HB estimation, by looking at the alpha file I describe whether they are significant or not.

However,I got confused whether it is the right approach or I should  compare their preferences on attributes based on their individualism, e.g individualist vs. collectivist respondents. I read that comparison of utilities or beta parameters for 2 MNL models is inappropriate. Swait and Louviere (1993) suggests a test for such situation. But I dont know how to apply it to my ACBC survey (no BYO, only choice task tournament).

I would be so glad if you could give me some advise about the right approach and how can I apply Swait-Louviere test, as I am really stuck at the moment, I could not find how to apply it. Is it possible with Sawtooth SSI ?


Any help would be great. Thank you !

This is  the article:
SWAIT, J. and LOUVIERE, J., 1993. The role of the scale parameter in the estimation and comparison of multinomial logit models.
asked Dec 12, 2014 by Mir (400 points)
retagged Dec 12, 2014 by Walter Williams

1 Answer

0 votes
Yes, we are familiar with that article and two of us here (Keith Chrzan and myself) sometimes use the Swait-Louviere test when comparing the results from two aggregate logit models, for two different groups of respondents.  

While it would be possible to do this with ACBC data, it would be very tricky to do, since you would have to deal with the complex organization of the ACBC choice data files, which involve different layouts for the BYO, Screener Section, and Choice tournaments.  Because you would need to be multiplying the design matrix of one group of respondents to try to match the scale factor to be able to compare to a different group of respondents, you'd need to then effects-code or dummy-code the raw choice data files for ACBC.  This would be a super laborious process.

In my opinion, the two approaches (one Bayesian and one Frequentist) I have been encouraging you to use for multiple posts now  would be easier to conduct and probably even better:  1) use of covariates in HB estimation, then examining the differences between groups of respondents on the part-worth parameters via the alpha file of population mean estimates, 2) use of the normalized part-worth utilities (the point estimates) per individual, via F-test or T-test on the part-worth parameters or on the attribute importance scores, comparing one group to another.

If you are using a continuous variable (such as your individualism scores), you can run a regression analysis with the individualism as independent variable and the normalized part-worth utilities (or the importance score to summarize the weight of attributes) from HB as dependent variables.  The t-value for the beta weight would give you your statistical test.

If you are just going to use the individualism score to divide the respondents into two groups, then you could compare respondents using F-Test or T-Test using standard cross-tabulation software statistical testing.
answered Dec 12, 2014 by Bryan Orme Platinum Sawtooth Software, Inc. (148,140 points)
covariate estimation (acbc)
...