Have an idea?

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

Interaction or not with CVA study?


I have a CVA study about "juice taste". There are three attributes, with "Sugar Level", "Stable Level" and "Pure Juice %" respectively. Each attribute has three levels. I generate the full-profile concept and ask respondents to rate each  concept using 10-point scale.

Now, I am at the modelling phase, and I wonder if I need to consider interaction among each attribute. Because I think for the juice, it is not adequate to examine only main effects, but also consider their interactions.

I have some questions about it, and it would be much appreciated if someone helps me.

1. I know it could be done two-way attributes interaction in CBC/HB, but not sure whether it could do three-way attributes interactions.

2. Can I run CVA HB in CBC/HB software, or I have to use another software? If I can run this study in CBC/HB, how to format the data file?

3. If it is not suitable to estimate utilities by HB, can I also apply interaction using OLS regression or Logit method?

4. Which software I could use to analyse the data except CBC/HB? (SMRT software? But I cannot find interaction option in it)

Thanks everybody for helping me in advance.
asked Jan 30, 2012 by ericdee Bronze (1,805 points)
retagged Sep 13, 2012 by Walter Williams

2 Answers

+3 votes
CBC/HB will not work for this situation.  It utilizes logit-based data, while CVA produces ratings-based data.  CBC/HB also does not handle three-way interactions.

SSI Web contains the latest version of CVA/HB, but with so much data at the individual level it does not estimate interaction effects.  HB-Reg can also estimate the data.

To estimate an interaction, you can combine two or all three attributes into a single attribute enumerating all possible combinations, and estimate using that.  You can then see whether there was an improvement.  HB-Reg may be the best tool to be able to manage the different codings.  To use HB-Reg for this problem, you will want to dummy-code your attributes remembering to code n-1 columns (n being the number of levels).
answered Jan 30, 2012 by Walter Williams Gold Sawtooth Software, Inc. (20,005 points)
Indeed, if this user is good at performing dummy-coding and understands how to code up a CVA estimation in HB-Reg (the design matrix, including design matrix for possible interaction effects), then HB-Reg is a possibility.  But, then additional work will also be required to get betas out of HB-Reg and prepared for either SMRT simulator or our Online Simulator.

The CVA/HB approach, and recasting the study with a collapsed attribute by creating a copy of the SSI Web study, and exporting the design + modifying the design + importing the modified design, will be easier.
+2 votes
It is very likely that you are stuck, and will not be able to estimate interactions with CVA.

Unless you plan ahead in CVA in the design phase of the study, you will probably not have enough combinations of products shown to respondents to support estimates of interaction effects.  I'm assuming you used the approach of 1 questionnaire version in CVA?  But, if you used multiple questionnaire versions in CVA (a feature of the newest CVA version), you might be able to do this, but it will require CVA/HB estimation (and re-organization of the design file, as described below).  Yes, I said CVA/HB estimation.  CVA has an HB module ($1K more to purchase).

The way to do interaction effects in CVA PRIOR to fielding the study, is to plan ahead of time  which 2 attributes will be involved in the interaction effect (I know, harder said than done).  Then, instead of specifying these as two attributes with 3 levels each, you should specify these as a single collapsed attribute with 9 levels.

Then, during estimation (we prefer CVA/HB rather than CVA + default OLS), you can estimate as all 9 levels (perhaps with additional constraints between selected levels within those 9 that have known order).  You get one attribute with 9 utilities, not 3 attributes each with 3 utilities.

Then, if after fielding the study you don't believe that there is an interaction effect that is strong between those two attributes...with additional work you can re-organize the design file (saving a copy of the original study to a new study to work in) and separate the attributes back to two 3-level attributes for estimation.  This saves degrees of freedom during the estimation and is a more parsimonious model.

Of course, CBC software makes the estimation of ALL potential 2-way interaction effects possible because it typically manages dozens or hundreds of versions of the questionnaire design across people.  This means lots of combinations of product concepts have been seen across the sample.  Via logit, latent class, or CBC/HB estimation, interaction effects are possible in CBC.

But, in CBC, if you only field 1 questionnaire version (not typical practice), you'd probably be in the same bad situation with respect to interaction effects as a 1-version CVA study.

BUT, if you use the newest CVA's capability of fielding 10 versions of the CVA questionnaire, then you are probably in very good position to estimate an interaction between any two 3-level attributes in your study.  Again, reorganization of the design file to collapse 2 attributes into 1 with 9 levels...plus CVA/HB estimation is needed.

Please call in to tech support if any of this is unclear, or if you need more details regarding how to save a copy of your study that recodes the attributes as one collapsed attribute rather than 2.
answered Jan 30, 2012 by Bryan Orme Platinum Sawtooth Software, Inc. (174,440 points)