I’m estimating an MNL and RPL model for a DCE that includes two policy alternatives and a “neither” (opt-out) option per choice set. Each alternative—including the opt-out—is represented as a row in long format.
For the opt-out alternative, I’ve set all attribute values to 0 (both continuous and categorical). These zero values do not correspond to actual design levels, but are reserved exclusively to indicate the opt-out. Below is an excerpt of the data structure for one respondent:
Given this attribute coding, is it necessary (or advisable) to include the explicit ASC variable for the opt-out alternative? Or is the opt-out already sufficiently modeled through its zero-coded attributes alone?
My goal is to avoid overparameterizing the model or introducing multicollinearity/redundancy.
Any clarification or best-practice advice would be appreciated!
You indeed need an ASC for an opt-out alternative. An opt-out alternative is a different type of alternative and hence requires a constant. So your data in long format (used by Nlogit and some other software packages) looks good.
You would immediately know if you overspecify the model or introduce multicollinearity because your model estimation would fail.
I’ve been running both conventional and hybrid MXL models with ASC included in my estimations. However, I recently came across a couple of papers with a similar setup to that of mine, published in Ecological Economics and Energy Economics, that omitted ASC from their models. They argued that ASC was effectively captured through the way they defined their attributes, which led me to second-guess my own approach.
Additionally, including ASC in my model results in an MWTP estimate that’s actually larger than the highest level of my cost attribute; this further raised my concern about potentially over-specifying the model. I am looking into interacting the ASC variable with observed sociodemographic variables to see if I can explain the high ASC standard deviation value.
In any case, I really appreciate your reassurance!
Including an ASC or including a dummy variable that represents the label of the alternative is the same thing. Perhaps that is what the authors did in that paper, not sure. Otherwise, ASCs should always be added during model estimation.
Your ASC expresses the preference between choosing and not choosing (the opt-out). This could indeed be large effect, because the vast majority of the decision makers do (not) like the opt-out alternative. WTP calculations do not include the ASCs.