Have an idea?

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

Covarates in max diff HB estimation

I am considering using covariates in an upcoming max diff study (to understand segment differences in max diff item preferences) – that will use Sawtooth software (SSI Web 8.3.4) for HB utility estimation. I have conducted several previous max diff studies involving HB utility estimation BUT have never used covariates. I’ve read the relevant CBC/HB covariate documentation – but would like to verify that the points below are correct since I am new to max diff score estimation with covariates:

(1) Covariates can be categorical OR continuous. When preparing the covariate CSV file, the covariates on their original scale can be used (for example, categorical covariate with 3 categories would be coded 1/2/3 and continuous covariates such as “planned spend on next PC purchase” would remain on its original dollar/currency scale). Is all of this correct?

(2) To estimate max diff HB utilities – you follow all of the same steps you normally would in SSI Web 8.3.4 – the only exceptions being (a) importing the CSV file containing the covariates into SSI Web using the merged variables import in the Data Manager (b) telling the software what variables to use as the covariates. Is all of this correct?

(3) The final max diff scores incorporating the covariate information that should be used for further analysis and client reporting are contained exactly where they would be if not using covariates: in the Studyname_scores_hb_report.xlsx file. Both the Raw and 0-100 rescaled max diff scores in this file will incorporate the covariate information. Is all of this correct?
asked Oct 30, 2014 by anonymous

1 Answer

0 votes
You've got this right.  The output will be in the same layout as when running HB without covariates.  Additional things to consider:

1.  Too many covariates can seriously slow down the estimation.  Too many depends on the data set and the machine's memory...but a good rule of thumb is to try to use covariates that would be about 10 or fewer columns (after dummy-coding or continuous variable coding) in the design matrix.

2.  Well-chosen covariates that correlate with choice behavior (on the MaxDiff items) will lead to more heterogeneity reflected in the item scores.
answered Oct 30, 2014 by Bryan Orme Platinum Sawtooth Software, Inc. (180,515 points)
Thank you for the helpful information