Have an idea?

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

How to interpret multiple covariates in a HB run?

Hi Support-Team,

Actually I have two questions:

1)    I read your technical paper about including covariates but I am still not sure how to interpret the alpha-file with more than one continuous covariate. I burn the first 10.000 draws, calculate the mean of the remaining draws for the intercept, the covariate environmentalism and Status consumption and each level. I also check the face validity of the covariates. Am I right up to this point?

Is it right to say interpret each covariate isolated, as if the other one is not existing?
Let’s say the value for Level A and covariate Environmentalism is -0,6 and for the covariate status consumption I have +0,5. Is it right to say something like that: ”If a respondent has a higher degree of environmentalism, the part-worths of level A decreases for that respondent..” That would the same interpretation as in your paper about covariates. Or do I have to take the Status Consumption into account?

2)    Is it possible to deal with seven point likert scales as continuous variables, when including them in my HB/CBC run?
Thanks for your help!

Best regards,
asked Jul 7, 2015 by Martin

1 Answer

+1 vote
Covariates should (like any set of independent variables in a regression) have enough independence (lack of multicolinearity) to allow their betas to be interpreted independently of the others.  If you suspect an interaction among your covariates, then you should code the covariates as an interaction effect for use in the covariates design matrix.

You can use likert scales as covariates in CBC/HB runs.  I'd recommend somehow normalizing them to have a mean of zero and constant variance (such as 1.0) across different variables.
answered Jul 7, 2015 by Bryan Orme Platinum Sawtooth Software, Inc. (131,390 points)
Hello Bryan, many thanks for your fast answer and your help!
I now normalized my covariates, which were a result of the Likert Scale. I also checked them for multicolinearity: they are independent!

But I have one further question:
I want to include one more covariate, which I received from an open-end question. The question was, how many cars the respondent owns. The answers range from 0 to 15.

Would you recommend to normalize this data as well?

Best regards, Martin
If you want to be able to interpret the magnitude ("Importance") of the covariate beta effect across covariates, then you should normalize it.  Otherwise, you should be able to just use the number of cars question as-is.