Have an idea?

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

Setting up relationship between attributes, conditional relationship?

Hi there,

In my CBC questions, there are two attributes, one is an amount of money, and another is time to pay back the money. I want to present two alternatives like this, for example, in first alternative money is £500 and time is 4 months, then in the second alternative, if money is higher than £500, time should longer than 4 months(i.e. £600, 5 months), if money is less than £500 time should shorter than 4 months(i.e. £400, 2 month). How should I set this in CBC design?

Many thanks
asked Dec 19, 2017 by hanjin (240 points)

1 Answer

0 votes
This is a complicated thing to do, but might be possible.  Two approaches come to my mind: cross-concept prohibitions and what is called the "Premium Pricing" approach.

The first thing to do is to create your experiment without any prohibitions and use the Test Design feature of the software to create robotic respondents who answer randomly, then to estimate aggregate logit effects and report the standard errors (the precision) of the attribute levels.  You are hoping for standard errors for each of your attribute levels of 0.05 or less.  This is your standard.  And, when you try to constrain your design to follow the patterns of more realistic tradeoffs you are trying to implement, you'll want to compare the results to this "no-prohibitions" standard.

If you have our advanced design module for CBC software, there is an advanced feature in the prohibitions area that allows you to input cross-concept prohibitions.  That is to say, you can prohibit a combination of attribute levels from occurring if another combination of attribute levels occurs for a different concept within the SAME choice task.  Try this, then run the Test Design feature again with the same number of robotic respondents as before.  Compare the standard errors.  You should hope that they don't increase very much.  And, you should hope that they are all still under 0.05 for each attribute level.

Sometimes when you are trying to implement so many prohibitions, the Balanced Overlap, Complete Enumeration, and Shortcut design methods cannot find a solution and report that there are "too many prohibitions."  This is a sign that you are in a danger zone.  But, you can push forward investigating whether it might be possible to do these prohibitions using the "Random" design approach.  Again, test the design and examine your standard errors carefully.

If this doesn't work, then another approach to investigate that often works better is the "Premium Pricing" approach, which is not documented in our software help.  It is an advanced strategy that has been taught at our Turbo CBC seminars by a consultant named David Lyon, from Aurora Market Modeling.  We have described the Premium Pricing approach in our new book, "Becoming an Expert in Conjoint Analysis" in chapter 3.

If you try the premium pricing approach, you should again create robotic dummy respondents and create an aggregate logit model to test the precision of your effects.  

Even though I describe using aggregate logit to test your models, you will likely do the final utility estimation using something like HB.
answered Dec 19, 2017 by Bryan Orme Platinum Sawtooth Software, Inc. (132,290 points)
Thank you for your detail explanation. I tried the prohibition approach, but it doesn't work, even I have set the prohibition, still, some unexpected combinations showed when I tested it. Is this situation can be solved?

Related to the second approach, I read about premium pricing in chapter 3. As the example shown in the book, they were talking the situation only with two alternatives(which is small and large package price). However, in my case, I have four levels in each attribute (attribute 1 payback amount:£200,£300,£400, £500; attribute 1 repaying time : 1 month, 3 months, 4 months, 5 months). In this case, what should I do?
Regarding the cross-concept prohibitions, if you are having a hard time getting the software to do those for you, I'd recommend you contact our technical support team at support@sawtoothsoftware.com.   Perhaps you are making a simple mistake.  Or, perhaps there is something wrong with our software.  It would be good to know.

Sometimes too many prohibitions are being required, so the experimental design algorithm cannot reconcile the requirements together with its goals.  In that case, sometimes selecting the "random" approach rather than the "balanced overlap" approach could work.  But, this is a red flag, and perhaps another approach could be more statistically efficient.

The "premium pricing" example, as introduced in chapter 3 of our advanced book, can potentially be a way to resolve your problem.  Extending beyond 2 attributes can be done, and it gets tricky fast.  David Lyon of Aurora Market Modeling is the most expert person I know regarding that particular approach.

Another approach is to use Sawtooth Software's experimental design tools to generate an initial design without prohibitions; to export that design to .CSV file , and to modify manually the design in Excel to impose the prohibitions you want upon the design.  Then, re-import the design and test it with robotic respondents to ensure that it can still estimate the parameters of interest with good enough precision.
Thank you Bryan. I will contact your technical support team about the prohibition.  
About the ''premium pricing'', I have read your book, but still don't know how to do it in sawtooth, do you know where can I find more detail information or tutorial about this?

Thank you.
Write me at bryan"at" sawtoothsoftware.com and I'll email you a Word document that tells you the steps for doing the premium price example that we describe in the book within Lighthouse Studio software.