[Limdep Nlogit List] How are attribute nonattendence models estimated in NLOGIT?

[Richard Turner]

Greetings,

I have fielded a choice study and I have information on which attributes
respondents explicitly ignored when making choices. After doing some
research, I found that I could code each attribute that was explicitly
ignored as -888 in NLOGIT that induces NLOGIT to switch to a "attribute
nonattendence model."

To my knowledge the attribute nonattendence model explicitly fixes
coefficients at the individual level to 0. E.g. if the utility function was:

U_ij = b_1*COLOR_ij + b_2*Price_ij


and the individual ignored color, then the utility function would become

U_ij = 0*COLOR_ij + b_2*Price_ij

Therefore, the sum of the utility that would enter the likelihood function
would only consist of b_2*Price_ij.

My questions are:

   1.  It seems that modeling these nonattended attributes using the method
   above is equivalent to modeling these attributes as *missing at random*.
   I thought that if data that was *missing at random* was modeled in this
   way, it would produce biased estimates. How can this method then not
   produce inconsistent and biased parameter estimates?
   2. Is it possible to estimate *mixed-logit attribute nonattendence
   models* in NLOGIT?
   3. Does anyone know of resources that give more of an explanation of
   these models, e.g. why they would produce consistent and unbiased parameter
   estimates?



______________________________________________

*Richard Turner*

*iMarketResearch*
*Easier, Faster, Better Advanced Market Research*
www.imarketresearch.com

Address: 12 Penns Trail, Suite A, Newtown, PA 18940