Hello, I use Choice Platform for Discrete Choice experiments and meet difficulties to implement "None of these options". Anyone knows how to handle it?
For example, I run simple case with only two attributes - brand and price.
- I add one more choice alternative "None of these options" in profiles table .
- Since nor price or brand is relevant to this option, I assign "None" level for the brand and price attributes for this choice.
- Respectively, "None" attribute has level "0" for all normal choices and "1" for "None" choice.
The issue is that in such setup JMP gives me calculation error. All utilities become zeros and errors become huge. If I replace "None" price with some other values from the levels what I have for normal choices, it works... but it's incorrect since the "None" option impact utility for given price level.
This is strange because "None of these options" is a must in DC experiments. So pity that JMP guides don't cover this topic.
So, what to do?
You can use a No Choice in JMP 12, though it takes a little work with the data table, and the best way to do it depends on you data and how you normally model the No Choice option (there are different schools of thought on that). I'm happy to work with you. You can message me through the forum or e-mail me directly for more information.
In JMP 13, we will be including an option to model "no choice" explicitly, and I'd love to talk with you about how you use this feature and the best way to report results from it.
JMP Research Statistician Developer
Hello, Melinda, thank you for your answer, love to see it here
I use separate tables for choice profiles and for actual choice run.
Within this setup, I can use either unique id for every choice or use grouping by choice set with sequential number of choice. It works both.
So far I have to totally delete "None" options and treat all answered choice sets with chosen "None" option as unanswered.
This is generally OK and gives more or less good information about utilities for real alternatives.
However, I would like to see the total utility for "None".
The logic behind this would be treating the "None" as "Other", so utility for "None" should represent preference for all other alternatives that was not shown in experiment. As soon as I know, this approach is the most common.
I'll cc you by e-mail but I think your explanations would be fruitful for the community.
Melinda answered me by e-mail, so I post solution here.
The idea is to re-code nominal variables as binary dummy variables, set their type as ordinal and conduct analysis by them, not by original nominal variables. One of level within each attribute should be leaved out as reference baseline with utility level=0.
"No choice" should be coded similarly - as 0 for all normal choices and as 1 for "no choice" choice.
As result, it becomes possible to calculate utility for "no choice" with zero baseline.
It works great, thank you, Melinda! Moreover, as an aside, it becomes possible to measure likelihood ratio test significance for each level, not for all levels combined, this is very useful. The only drawback is a bit more work to construct a model (more clicks), especially for interactions.
I added screenshot to illustrate changes in standard sample to accommodate "No choice" option.
And I've attached are a couple of sample data tables that you can use to try it at home!
Thanks for posting the information for us. I like your idea of giving the "no choice" option a price, indicating that you think the consumer would be buying another brand with a lower price. How you choose to code the attributes of the "none" is up to you--just be careful when you interpret the results.
For others who are interested in this topic: Using an indicator variable for the "none" response can cause bias in your parameter estimates. This paper discusses an alternative model for the No Choice option using a nested multinomial logit instead of an indicator variable.
Vermeulen, B., Goos, P., & Vandebroek, M. (2008). Models and optimal designs for conjoint choice experiments including a no-choice option. International Journal of Research in Marketing, 25(2), 94–103. doi:10.1016/j.ijresmar.2007.12.004
I'm currently working on some comparisons between the indicator and the nested logit. If others are working on this problem, I'd love to hear from you.
Thanks for sharing this neat work around for the none option for versions of JMP older than 13!
I have used the files that you shared regarding the Pizza profiles and responses with the none option included, which is much appreciated.
However, now that I have had a play around with them I have come across a problem - for me at least and perhaps for others too.
If you change which indicator variables (attribute levels) are include when you run the model the value of both the utility and X2 change fairly dramatically.
So how do you interpret the value of "none"?
For example if you include "Mozzarella", "Pepperoni", and "Thick", the Utility value of none is -1.68 and X2 is 15.025, which is great and can be interpreted if this value doesn't change when you change attribute levels, however if you include a different combination of attribute levels lets say the inverse "Jack", "None" and "Thin" it takes on a completely different set of values for none Utility is now -3.29 and X2 is 76.703. Yet the AICc and BIC remain constant as does the -2*LL, and as expected the inverse values for the utility of each of the levels that were included compared to the first run.
Intuitively this seems wrong... as the first combination of levels of the alternative are much more favorable than the second yet the utility for none gets less for the less favored combination, yet the X2 increases, suggesting it is more weighted in the analysis?
I looked through the results a number of different ways and exported a data table from the profiler which I think highlights the problem (perhaps)...
The attribute parameters appear to being estimated twice once with none and once without? Shouldn't there only be 9 profile estimates, the 8 alternatives plus none/no choice (coded as 0 0 0 1)? If I am wrong could you please explain how I should interpret the none in your examples?
Thanks so much this is very helpful!!
Hi, Jasha. The problem in your output data table is that you set "No Choice" = 1, and then you set attribute levels for the other variables.
The "No Choice" column isn't a comparison. The Choice platform treats it like another factor. When you set up your profiles table, you should have a "No Choice" product that has 0's for everything except the No Choice factor: All of the other products should have 0's for the No Choice factor. Like this:
Notice how the "No Choice" column is all 0's except for the last row in the table? Notice also that that row has all 0's for the other factor? The last row in the table is the "None" or "Competing" product.
Hope this helps!
Hi Melinda, many thanks for your response.
I definitely understand what you are saying here, however, my profile table is set-up as you have suggested with the "No Choice" column as all 0's except for the last row in the table, this is because I have used the files you attached above a few months ago.
The resulting output table from the profiler is generated when I use your original profiles table with the none option set as an dummy variable, and the toppings each as dummy coded variables, and I run the analysis as pictured.
And I get the following basic output table:
If I change the combination of toppings or crust type estimated along with no choice, the value of no choice changes in terms of X2 and utility where as the toppings and crust (because they are binary in this case and dummy coded) just change sign i.e here I have changed "Jack" for "Mozzarella" only and as you can see the no choice option has now changed values:
So I'm not sure how to correctly interpret no choice? I am sure perhaps I'm overlooking something, but I can't see what that is?
Once again thanks for answering and I hope you can help with this followup question!
Just like in the profile table, settings in the Utility Profiler that include No Choice =1 and other factors also set = 1 are not interpretable. JMP doesn't know that it's impossible to choose "none" and also choose Pepperoni=1, so when you use "Output Grid Table" (I assume that's how you got the table in your initial question) JMP shows all possible combinations of factor settings.
So, the only rows that are interpretable in your output table are Row 1, and Rows 10-16.