Have an idea?

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

Coding and specifying a regular CBC exercise in MBC

I fielded an experiment with 2 options + 1 none. For various reasons we needed to analyze the data in MBC and it would be great to get some feedback on my process below.

To get the data into MBC, each row contains the attributes' expressions for the 1st option, than the same attributes' expressions for the 2nd option and then the DV/choice, correct?

When I specify the model the 1st set of variables (attributes of option 1) affects only choice 1 and the 2nd set only choice 2. (None=offstate)

To get the same parameter effect for a specific attribute across both options, I have to collapse the variables reflecting this attribute for option 1 and 2, correct?

From what I see in the preview it looks like this should yield the same specification as the CBC .csv format that I get when I transforming a .cho file or exporting straight from Lighthouse. Is that the case or is there any other "buttons I have to click" to get the same output?
related to an answer for: counts vs utility part worths
asked Jan 8 by alex.wendland Bronze (2,025 points)

1 Answer

+2 votes
 
Best answer
What you describe is one of the examples we went over in the MBC training, which I believe once upon a time you attended.

And, you are remembering correctly: for main effects of each attribute (rather than alternative-specific effects) you need to use the "collapse variables" functionality to collapse (make generic) the attribute effects across the alternatives.

Once you do that, you'll get the same result as regular CBC, except that MBC uses dummy coding and CBC uses effects coding (such that the first level is constrained to zero in the MBC case, but the average of the utilities within the attribute is constrained to zero in the CBC case).
answered Jan 8 by Bryan Orme Platinum Sawtooth Software, Inc. (150,315 points)
selected Jan 9 by alex.wendland
Thanks for confirming Bryan. I wasn't sure anymore in light of the related query regarding the discrepancy between counts and utility estimates. Would you have any further ideas for investigating this "anomaly" in the linked post? Should I pick this up with Walt maybe? I think I really need advise on the interpretation of this contradiction for the far-reaching implications for our client.
Well, when I was in your position many years ago I was very paranoid and skeptical of any software program, so I often would run the data through multiple different programs provided by different software providers--especially if something seemed fishy.  

And, did you just try to isolate this variable only as a single independent variable, to see if the MNL coefficients better aligned with counts?  Perhaps the issue is due to a problem in your experimental design.

Do you have another MNL program at your or at your colleagues' disposal there in which you can format the data and submit this MNL run?
...