Have an idea?

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

Checking Covariates using pencil-and-paper-data in Version 8.3.2

To whom it may concern,

We would like to investigate the effect of a few covariates (such as gender and expertise) on the part-worth utility values. We have read the Technical Paper on "Application of Covariates Within Sawtooth Software's CBC/HB Program: Theory and Practical Example", but are having trouble viewing the results as we only get an hb_alpha.csv file. Is there any way we can "ask" the program to give us a summary of these results (using Sawtooth version 8.3.2) to facilitate interpretation?

Also, is it possible to look at the effect of multiple covariates simultaneously? Is there a general rule as to the number of observations per covariate (e.g. 10 observations for every new covariate)?

While thanking you beforehand for your help, we look forward to hearing from you!

Nadine and Laura
asked Jun 6, 2014 by anonymous

1 Answer

0 votes
We don't provide automatic processing or summary of the hb_alpha.csv file.  But, you can do it easily in Excel:

1.  Ignore the first 10,000 rows of the file (you may delete them)...or as many iterations as you feel should be thrown away because they don't represent the draws after convergence has been obtained.

2.  Then, just take the means of the remaining columns.  Those are the mean estimates of the covariate weights (they are to be interpreted like regression weights).  

You may use multiple covariates at the same time within CBC/HB.  I'm not aware of a general rule regarding the number of observations per covariate required.  But, you will find under some data or system conditions that the software can slow down considerably if more than about 10 to 15 "columns" in the coded covariate design matrix are used.  So, that would be 10 binary covariates or a single 11-category nominal covariate.
answered Jun 6, 2014 by Bryan Orme Platinum Sawtooth Software, Inc. (199,115 points)