Pilot study & Bayesian D-efficient design

[psalazar]

Dear Michiel,

I run a pilot study using the following attributes and levels:
-yol (years of life), three levels (0.5, 1, 3)
-qol (quality of life), three levels (20, 40, 60)
-exp (patient experience), three levels (0.poor, 1.fair, 2.good)
-size (size population), three levels (50, 5000, 10000)
-equ (equity), three levels (25,50,75)
-cost, three levels(20,40,60)

In the pilot study, I ran two separate D-efficient designs, as I am interested in how preferences change when ‘cost’ is not one of the attributes. In each design, I used priors close to zero, and used a partial profile design to have (in all choice sets) overlap in two attributes:
- Design 1: we treated all six attributes as numerical, except for “exp” which was treated as categorical
- Design 2: we included the first 5 attributes from the list (i.e. we did not include “cost”), and treated “yol”, “qol” and “exp” as categorical. “size” and “equ” were treated as numerical. I did this to make sure we have some choice sets with overlap in “exp” (I followed here your advice discussed in the Ngene forum ‘Candidate set & overlapping’)

I now have the pilot data collected and was wondering whether it would be okay to use this data, estimate the MNL (separately for Design 1 and Design 2) with all attributes as categorical, and then use the parameter estimates and the s.e. as priors into the Bayesian D-efficient design and treat all as categorical in Ngene?

I have heard some scholars arguing that when doing the design optimisation we should treat all attributes as categorical, and separate the experimental design stage from the estimation one. I wonder what would be your take on this?

In addition: Although it might be a bit late, I am now considering adding some interactions between attributes. If I am adding that into the Bayesian D-efficient design, to obtain the design for the actual study, should I assume priors close to zero? And what happens if I do not know the direction of the interaction terms?

Many thanks!

Best wishes,

Pamela.

[Michiel Bliemer]

Yes you can estimate an MNL model and then use the parameters are priors for generating a Bayesian D-efficient design.

I would not say that you SHOULD treat all attributes as categorical during the design phase, I would say that you COULD treat all attributes as categorical during the design phase. You can then estimate then as categorical or linear attributes during model estimation. Treating everything as categorical in the design phase may make the design size (=the number of choice tasks in the design) larger than perhaps necessary, but it does not hurt, it is a fine strategy.

Most people do not include interactions during the design phase, but if you do, you are at least guaranteed that you can estimate them (but in most cases you can anyway, unless your design size is very small). I typically set the priors of interactions to zero, only when I think that they are important (=have a large contribution to utility) then I may consider non-zero priors for interaction effects.

Michiel

[psalazar]

Many thanks, Professor.

One last question about the interactions: I am interested in the statistical significance of the interactions, rather than in the functional form of the parameter estimate of the interaction. Can I, therefore, add an interaction between, say, "cost" and "exp", with "exp" as categorical and "cost" as continuous, while having cost's main effect as categorical (and "exp" as categorical, as well)? In the Ngene manual, it says "Dummy coding of an attribute level in an interaction does not require that that attribute's main effect be dummy or effects coded", but I just want to double-check if it is still the case when I treat the variable as categorical in the main effects, but as continuous in the interaction.

Many thanks!

Best wishes,

Pamela.

[Michiel Bliemer]

Yes in estimation you can treat a variable as categorical in the main effect and as numerical in the interaction effect. I believe that that Ngerne also allows this in the utility function (although it is unusual and have never tried myself).

Michiel

[psalazar]

Thank you, professor!

Best wishes,

Pamela.