Have an idea?

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

ASC manipulations in CBC-like MBC

Hi All,
I have a MBC study, where a dependent variable has several choice categories plus "None" in it (off-state).
In order to handle "None" response estimation I brought ASC into the analysis. But I am getting ASC values for all response options in Dependent variable, not just for None category in it (which is constrained to Zero due to its off-state coding).
As I understand, these ASC values represent observed but unexplained residue  or predisposition to select each of choice options shown.

I am thinking that negative average of these ASCs might be a good proxy to "None" utility.
 Why?

Share=exp(A+ASC1) / (exp(A+ASC1)+exp(B+ASC2)+exp(C+ASC3)+exp(None+ASC4))

Since ASC4=0 and None has total zero utility, also assuming there is not that much difference between ASC1, ASC2 and ASC3  ~= ASC

Share=exp(A+ASC) / (exp(A+ASC)+exp(B+ASC)+exp(C+ASC)+exp(0+0))

or

Share=exp(A)*exp(ASC) / (exp(A)*exp(ASC)+exp(B)*exp(ASC)+exp(C)*exp(ASC)+exp(0+0))

dividing both nominator and denominator by exp(ASC) we would get

Share=exp(A) / (exp(A)+exp(B)+exp(C)+exp(0)/exp(ASC))

But exp(0)/exp(ASC) = exp(0-ASC), so

Share=exp(A) / (exp(A)+exp(B)+exp(C)+exp(-ASC))

In order words my reversed sign average ASC took place of "None" response option utility.

Is it conceptually right?


NOTE:  I understand I would get exactly the same share without all these tricks if I used straight ASC1, ASC2 and ASC3.  But I am trying to model choice tasks with different number of options, not just 3.
 So I am not that interested these specific  ASC1, ASC2 and ASC3 linked to order of options or whatever unidentified factors or biases as much as I am interested in proper None estimate.

NOTE:
I didn't code and structure data that way that would allow my dependent variable to be 0-1 only. This would have accommodated straightforward "None" estimation for sure.  But I felt that merging other predictors in the model the way they were actually shown in choice task would allow better estimate of these merged predictors. This required having data and dependent variable structured the way  I have.

Thanks
asked Nov 23, 2017 by furoley Bronze (545 points)
edited Nov 23, 2017 by furoley

Your answer

Please only use this to answer the original question. Otherwise please use comments.
Your name to display (optional):
Privacy: Your email address will only be used for sending these notifications.
Anti-spam verification:

To avoid this verification in future, please log in or register.
...