Have an idea?

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

What if preference order changes between raw values and zcd?

In one of my estimations, the order of attribute levels within one attribute (in terms of increasing part-worth utility values) changes when you switch from raw values to zero-centered differences. What does that mean for interpretation?

In general: How would you interpret the values of zero-centered differences (in contrast to raw utilities)?
asked Feb 14, 2017 by summand (385 points)
retagged Feb 14, 2017 by Walter Williams

1 Answer

+1 vote
One of the things to remember about part-worth utilities is that they are estimates, not exact "true" points.  There is a range of uncertainty surrounding each one.  We can compute a standard error (the standard deviation divided by the square root of the sample size) and then multiply the standard error by +/- 1.96 to obtain a 95% confidence interval.  We are 95% confident that the true utility falls within that range.  

My guess is the change in preference order you saw for levels within an attribute when comparing zero-centered diffs to raw utilities did not represent a statistically meaningful change in preference.  But, you could compute the 95% confidence intervals for either (or both) of the raw utilities or the zero-centered diffs to see if the 95% confidence intervals overlapped for those two levels that traded places.

Zero-centered diffs does a better job more equally weighting respondents than raw utilities, where some respondents (the most clean and certain ones) can have a larger weight in influencing the mean utilities.

The interpretation of the utilities follows the same rules.  Zero-centered diffs are just placed on a bigger scale.  The overall magnitude of the scale does not change your interpretation of the part-worth utilities.  You could multiply all the part-worth utilities by 2 or 100 or 1000 and your conclusions should not change.
answered Feb 15, 2017 by Bryan Orme Platinum Sawtooth Software, Inc. (186,865 points)