Have an idea?

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

CBC - partial profile

Thanks for the response Bryan. Sounds like there's still lots to learn.

If we decided the partial profile adaptation of the CBC choice task is appropriate, is there a suggestion on how many more questions to ask to make up for any insights "lost" due to fewer elements per task?
related to an answer for: Magic number of choice-task in CBC?
asked Mar 6, 2015 by WTW_cal Bronze (1,145 points)

2 Answers

+1 vote
Well, the simple answer is just to add as many tasks such that each level appears as many times as it would have under a full-profile study.

For example, imagine that you had a 8-attribute full-profile study that you were comfortable asking 10 tasks for to obtain enough information at the individual level.  If you decided to do a partial-profile approach showing just 4 of the 8 attributes in each task, then you know each task contains half has much information.  So, to compensate, you double the number of tasks to 20 to achieve the same amount of information as you had before.

Keith Chrzan has done research showing that the response error for partial-profile tasks is lower than for full-profile, so Keith would argue you don't need double as many tasks in the example above because the loss in sheer information for PP is counteracted by the reduction in respondent error.  What exactly the tradeoff is, however, I don't know.
answered Mar 6, 2015 by Bryan Orme Platinum Sawtooth Software, Inc. (132,290 points)
+1 vote
In empirical tests we consistently found that partial profile questions result in a fourfold reduction in response error relative to full profile questions.  As you note, however, we show fewer attributes per task and this results an increase in statistical error, typically around 50% but sometimes more and sometimes less (you can find out how much more or less for your study using the design test feature of the software and comparing the strength of design for the full and partial profile versions of your study).  

The two errors cancel one another to some extent but the net result when we have made empirical comparisons has almost always been a reduction in total error in the neighborhood of 50% with partial profile questions.
answered Mar 6, 2015 by Keith Chrzan Platinum Sawtooth Software, Inc. (50,900 points)