Have an idea?

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

CBC conjoint program question

I have an upcoming project that will include a CBC exercise.  My client asked the following and as I don't believe CBC can use constructed lists I wasn't sure if it could be done in one exercise.  If it was two different exercises then I wasn't sure that they could compare the two.  Any input you can provide based on what my client has asked would be helpful.

•    There is only the one attribute that should not be shown in the UK, so is it possible to program it as one design and make that attribute display conditional on region?
•    If that is possible, what impact does that have on the standard errors for the design and the attributes? How does that compare to the standard errors if we had two different designs for the US and the UK? I think we are targeting 200 completes in the UK and 300 in the US, but will have to confirm that.
asked Sep 14, 2019 by Jay Rutherford Gold (37,685 points)

1 Answer

0 votes
Jay,

If I'm understanding correctly, an entire attribute (with all its levels) is to be removed in the UK sample.  That means an entire row should disappear in the CBC question when showing to UK respondents; whereas US respondents see that additional row in the CBC questionnaire.

You could use conditional display or unverified perl IF to accomplish that.  However, it leads to some interesting challenges if you want to try to do this as a single CBC exercise (and single analysis).

UK respondents would still have it represented in their data file that they saw levels for that conditional attribute (1s, 2s, 3s etc. would be recorded in the data for levels of the missing attribute).  That means that the analysis would look at that design info for UK respondents and try to fit utilities to UK respondents as if they were reacting to levels of that missing attribute.  This would introduce a tiny bit of random noise to the utilities for the other attributes.  Plus, you'd have non-zero utilities in the utility file for that missing attribute based on draws from the population means (which included US respondents).  The way to correctly deal with this would be to post-process the data file to make it look like the UK respondents saw levels coded as 0 (missing) for the missing attribute, not 1, 2, 3, etc.

Creating two CBC exercises in the same study would allow you to delete the attribute from the design for the UK people.  That would mean you'd need to estimate utilities separately for the two populations and have two different simulators for the populations.  There would be two completely different experimental designs.

But, you could take the second route above and with a bit of fancy data processing on the .HBU file, you could combine the two groups of respondents and add utilities of 0 for the missing levels of the missing attribute for the UK respondents.  This allows you to keep all the respondents together in the same market simulator.

Regarding the standard errors, the standard errors for the non-conditional attribute are essentially unaffected.  But, the standard errors for the missing attribute would in reality be larger and achieve the precision that only the US respondents could support.  The UK-based respondents (if the data are processed correctly and the levels for the missing attribute are set to 0 in the experimental design for the UK respondents) do not contribute any information to making the conditional attribute any more precise than you would get from the US-only respondents.

Those are my thoughts.

--Bryan
answered Sep 16, 2019 by Bryan Orme Platinum Sawtooth Software, Inc. (177,015 points)
Oh, I should add that if you take the first route I described above which involves combining the two sets of respondents as a single datafile for analysis, you'd still end up with utilities for the UK respondents based on draws from the population mean (affected only by the US respondents for that missing attribute).

If in reality that missing attribute doesn't apply and should have no importance to the UK people, then the utilities should be set to 0 for the levels of that missing attribute that UK people didn't see.
Thanks for the feedback Bryan, that is helpful.  I will pass this on to my client to see how they'd like to proceed.  I'm just doing the programming on this and passing the data on to their analyst for them to set up their own simulator.

I've never had to manipulate any conjoint data as you suggest.  Would this be done in an .hbu file then?  If so, what would be the best way to correctly accomplish that to change the correct data to 0 without compromising the format of the file?
In one case I was suggesting manipulating the raw data file (often the .CSV or .CHO files) to insert 0s instead of levels 1, 2, 3, etc.  In the other case I was suggesting manipulating the estimated utility file (.hbu file) to change utilities to zeros for an attribute that is not relevant (should have no importance) to a respondent.

These files often can be manipulated often in Excel, then saved out as either .csv file or when it's a space-delimited file like the .hbut ifle, then you save out as text space delimited.
...