[Limdep Nlogit List] Effects Coding and Attribute Nonattendance using -888

[Nathan P. Kemper]

Hello,

I was hoping to get an opinion on how modeling the coding strategy used, effects or dummy coding, may relate to modeling attribute nonattendance (ANA) in Nlogit using the -888 coding for attributes ignored.  There are many examples in the literature where folks use both effects and dummy coded data when dealing with ANA.  However, recently I've run across a couple of article that make a case specifically against effects coding and ANA methods where a zero coefficient value is used for a respondent ignoring the attribute.

Here is the statement verbatim:  "Dummy coding is necessary for the use of the attribute nonattendance restrictions that assign zero on the coefficient values. Posing this zero restriction on a binary effect-coded variable {-1, 1} would not be equivalent to a zero weight in the utility function, but to a weight which is intermediate between absence and presence of the attribute. Dummy coding is less than ideal because of the potential confounding with the 'nobuy' alternative-specific constant, but in our case, this should be mitigated by the very low probability predicted for the no-buy option by all our models."

Seems to me that the point of effects coding is the midpoint of zero and the -1 and 1 indicate which direction the respondent's choice is from the base value of zero.  So I do not really understand the concern about "intermediate between absence and presence of the attribute" if using the -888 (or a zero) when the attribute is being ignored?  Zeros are used to indicate the no-buy option to indicate a zero level of all attributes when the no-buy is selected.  So to me it seems that this would work the same way when a respondent ignores an attribute and the -888 is used in place.   The path taken to zero seems unimportant; changing a 1 or a -1 to a zero seems like the same thing to me and I don't know how zero being the midpoint of the two indicates that effects coding should not be used.

I've seen researchers using effects coding and accounting for inferred ANA (using both mixed logit models and latent class) and stated ANA where zero coefficient values are used for respondents ignoring attributes.  I'm hoping to get some guidance on this subject from the group.

Thank you,

Nathan Kemper