labelling alternatives

[medard kakuru]

hullo, am doing a study on preferences for sorghum grain. am specifically studying three products made out of the grain, that is, porridge, beer and bread. my question is whether I should make a design for each of the products. which is the best way to handle this? what complicates the matter is the fact that not all my potential respondents process all the three products mentioned.

thank you

[Michiel Bliemer]

This is more a general question rather than an Ngene question. I do not know the background of the study, the aim of the study, the type of respondents, nor have the knowledge in this domain, so I am afraid I cannot answer your question. In Ngene, it is possible to create a design using the model averaging approach in which you can specify different models. Each model may have a different subset of alternatives. Maybe this approach is suitable for you?

[medard kakuru]

hullo Michael,
I posted a question about labeling alternatives. you said it more of a general question. to be specific, the background of my study is thus: there is limited demand for sorghum products due to undesirable attributes in the current sorghum varieties. the aim of my study is to determine the grain attributes that consumers prefer. the major sorghum products I want to consider are porridge, bread and beer. my respondents will be consumers of any of these products. I hope this information is rich enough to enable me get the best advice. thank you.

[Michiel Bliemer]

I was actually trying to say that since this is an Ngene forum, I can answer questions on the use of Ngene but not provide general information on how to design stated choice experiments (although we sometime do here on this forum, I hope you understand we cannot answer all such more general questions). In your case, I think you have to decide on what is the choice you would like your consumers to make? Is it the choice between these products (labelled alternatives), or would you like to investigate attributes of each product and vary them (unlabelled alternatives) in order to capture trade-offs between these attributes. Do you really want to let consumers make a choice between porridge, bread and beer? (but perhaps I do not quite understand your study, I also do not know what sorghum is, I am merely a simple transport planner ).

Another option is to use scenarios (nesting), for example something like

U(productA) = b1 * product[0,1,2] + ... /
U(productB) = b1 * product[product] + ... /
U(none) = 0

where 0 is porridge, 1 is bread, and 2 is beer. So you state a certain scenario "Suppose you would like to purchase bread", and then show ProductA and Product B (both being bread) with varying levels. But perhaps this is not what you want, perhaps you want to be able to compute market shares for certain products, in which you will need labelled alternatives. This is up to you to decide based on experimental design theory and I am afraid I cannot help you with.

Regards,
Michiel

[medard kakuru]

Dear Michael,
thank you for the feedback. am trying all the options that you are giving. though you already pointed out that my questions are general, am gaining a lot from you. my interest is actually to investigate attributes of each product and vary them in order to capture trade-offs between these attributes.

secondly, one of the most important attributes for those products is colour which has 4 levels. am trying to use the dummy coding syntax as per example in the manual on page 123. I have two questions here. you started from an efficient design where one already has the priors for the colour levels. don't I need to specify the dummy effects right from generating orthogonal designs so as to obtain priors for each of the levels? in otherwards, how to I obtain priors for those levels if in the first orthogonal design (before the pilot study) I had not specified the utility model with those dummy effects?
secondly, to use the dummy coding syntax, I need to create l-1 parameters. am wondering why am seeing only one parameter (b2) for the attribute A instead of two parameters since A has three levels.

thank you,
regards,

medard

[Michiel Bliemer]

Orthogonal designs are generated under no assumptions on the model type nor parameter priors. You can specify dummy coding in the utility function, however, for orthogonal designs will not take such information into account, only efficient designs do. Clearly, you can also create an efficient orthogonal design. In case you do not know any prior values, you can use zeros.

On page 123 of the manual, it states:

b2.dummy[1.2|0.8] * A[0,1,2]

In other words, parameter b2 is dummy coded (with the last level, 2, the base level), where 1.2 is the prior for the first level, and 0.8 is the prior for the second level. Hence, there are two parameter priors specified for 3 levels.

Regards,
Michiel

[medard kakuru]

hullo Michael,

I try to get your explanation. my question still is how to obtain the two priors for the two levels if I had not provided for two parameters in the initial utility function. I thought that in order to obtain two priors, I must have made a provision for two parameter estimates in the initial utility function, which is the one expressed when generating an orthogonal design. hope am right when I say that an orthogonal design is the first step towards obtaining priors.

In case I decide to use zeros as my priors, wont the design be sub-optimal?

I will be greatful to be enlightened more about these.

[Michiel Bliemer]

An orthogonal design is optimal if the priors are set to zero. An efficient design is optimal if the priors are correct. Using an orthogonal design basically means setting all priors to zero. You get priors from a pilot study using parameters from previous studies or using a pilot study with an orthogonal design or whatever you choose to do. There is no best way of doing this. There have been several discussions on how to obtain priors here on the forum.