Have an idea?

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

multiple regression with utilities from CBC

Hey Forum

I hope it's ok to attach some more questions to this (old) tread as this affects my actual analysis a lot.

I have two problems when I perform multiple regression with my utility variables:
1) there is fairly no correlation between the independent and the dependent variable(s).
2) And (therefore?) I have very low adjusted R-square values, no matter what I do (use separate items, use factors calculated etc.)

Do you have any idea why this can happen? Speaking from the content point: it simply does not make sense not to have correlation between those items. An also, what I can do to fix this?

Thanks a lot!
asked Mar 19, 2019 by bs77 Bronze (790 points)

1 Answer

0 votes
Hi, Boris,

I have seen this on some of my data sets, but on others I see very nice correlations between my utilities and my rating scale variables.  

When I've seen this happen, it's often in studies where I suspect my rating scale questions have a lot of halo effect (in a lot of brand image studies, for example, there is very little BUT halo effect going on).  

Your case does sound extreme, however.  I don't now that I have an answer for you.  Do the conjoint utilities on their own make sense?  Do your other variables, on their own, make sense?  I take it your answer to both is "yes" or you wouldn't have asked the question, but it always pays to ask.
answered Mar 19, 2019 by Keith Chrzan Platinum Sawtooth Software, Inc. (86,950 points)
Hi Keith

Thanks for your quick answer!

Yes, thats the point. It both makes sense. The utilities do show interpretable values and so do the predictor variables. The factors are reasonable, built on solid factor loadings with variance explained above 50%.

Although it is the same sample I appears that the CBC does not fit to the rest of the data (or the other way around). I really can not explain this.

When I perform HB-Analysis I do have a Pct.Cert of about 0.6 which is low (this was already a topic between me and Bryan) but the best I can get without overfitting the model. Using co-variances did not make any sense, maybe also because of the lack of correlation. Interaction effects are not measured because of some design issues (ngene).

The predictors itself are nearly normal distributed. The dependent variables are not, but as I heard this seems not be be crucial (ok depends on the point of view but I don't think that this is crucial in my situation).

I checked the frequencies of the majority of the items (before factor analysis) and again it does make sense...

Unfortunately, I really don't have an answer for you.  I do not know why that is happening with your data.