does anyone know anything about the difference between labelled alternatives and unlabelled alternatives in terms of estimating and interpreting choice models? can we change the unlabelled alternatives to labelled alternatives?
I am a bit unclear about what you mean by "Labeled" versus "Unlabeled".
Does Labeled refer to all the choices that are known and available to all persons in the experiment (hence dataset), and Unlabeled refer to choices that may range accross individuals, but their availability is known only to them?
For instance in the classic commuting example, a person can choose to commute to work by 1)Bus, 3)Train, 4)Car. These are the "Labeled" alternatives. However, there are other options that may not be in the labeled choice set such as: walking, telecommuting, or quitting your job.
In this context, I would look at a Nested Logististic (or 2 stage model) where the first stage estimates the likelihood the person selected a labeled alternative. The second then estimates the likelihood of each specific alternative.
If the choices you present have two or more options, choice sets must be created. Choice sets could contain several alternatives defined by a set of attributes and attribute levels. The choices could also be binary or includes two ‘forced’ alternatives (i.e. alternative A or B). To answer your question on labeling alternatives, I think asking respondents to choose between two such alternatives would have an effect on your estimation and interpretation of the choice models. If you include opt-out options, you can use “I can’t choose” or “I am indifferent between the options” for instance.
If you have more than 2 alternatives, then you may want to think of the burden you could create for respondents if alternatives are not labelled. Rather crucial is the design that is used to create these alternatives whether labelled or not. So It is possible to label alternatives that are not labelled.
Let's start from unlabelled or generic alternatives. Say that the choice set contains discrete alternatives consisting of consumption bundles X, Y, ...,Z. The consumption bundles differ only in the amount of the various consumption components, i.e. their attributes. In general, a choice model with unlabelled or generic alternatives is one in which the alternatives differs only in the observed attributes (apart from the stochastic component). In terms of model specification, this case leads to the Conditional Logit without alternative-specific constatnts.
Turning to labelled alternatives, consider the choice between personal childcare (e.g. provided by family members), market childcare (e.g. provided by a baby sitter) and public chilcare (e.g. provided by a state kindergarden). Each of these alternative might be described in term of attributes e.g. cost, hours of service, distance from home etc. However, each of those alternatives has probably also other systematic (non random) features that are not observed but affect the attractivenes of the alternatives and make them different beyond the differences in the observed attributes.In terms of model specification, this case typically leads to a Conditional Logit with alternative-specific constants (possibly interacted with personal characteristics).
The extreme case of labelled alternatives is the one where no observed attributes are present, only the label distinguishes the alternatives. As far as the model specification is concerned, this leads to a Multinomial Logit with only alternative-specific constants (as before, interacted with personal characteristics).
Label of alternative is one of attribute of choice alternative. When we put a label on alternatives, we would know how the respondent react to the label, ceteris paribus. That's why you cannot change unlabelled alternatives in labelled after the person replied to the questionnaire. But some times you can use nested logit model with unlabeled choice set (look examples in K. Train's book)
Two main differences between estimating labelled and unlabelled alternatives:
- Alternative specific constants must always be used with labelled choice in order to obtain real market share of each alternative. While it is non sense to use alternatives specific constants with unlabelled alternatives.
- In labelled choice set some attributes are alternative-specific. For exemple the price of bus ticket is in general less expensive than the cost of travel by car. Or you can use number of flight connections only for air travel.
http://eml.berkeley.edu/books/choice2.html
Book Discrete Choice Methods With Simulation
Experiments where the alternatives are specified with a label (lets say 'car', 'bus', 'train' in case of choice of travel mode) are called labelled experiments.
On the other hand, the experiments where the alternatives are not specified with any label, and the attributes (travel time, travel cost, comfort) are generic are called unlabelled experiments. In these experiments decision maker has to opt for his choice (travel mode) on the basis of the levels associated with these attributes.
I recommend you to go through the book by Prof. David Hensher, John Rose, and William Greene. Book title is "Applied Choice Analysis - A Primer"
This book can make you fundamentally sound pertaining to Choice modelling and experimental design.
With Regards
Ashu
If I anwt to do an unlabelled choice model in R, I can use the mlogit package?, because I' been reading about thispackage, but I don'n find how to run these kind of model in R.
thank you
I agree with Ugo Colombino .
One example about labeled choices you can find in my paper about WTB and WTP for labels wines. Have a look into my publication. We use in our research artificial neural networks and its derivative for the analyses as alternative to a classic logit model.
- Badges
- Science topic
- Similar topics
- Philosophy
- Esthetics
- Interpretation