Pilot testing using an D efficient design

[Nancy Wang]

Dear Michiel,

I am conducting an experimental design with 6 attributes (4 with three levels and 2 with four levels; two of them are numerical levels). I now have a few questions I would like to consult you about. Note: I used two check constraints, so the modified M algorithm was employed.

1. I want to set three options (i.e., alt1, alt2, none). When defining the utility function, U(alt1) = b1*B, U(alt2) = b2*B. Should U(none) be set to 0 or represented by b_optout? I tried setting U(none) = 0/b_optout separately and found that the D-values from the two results differ.

2. One attribute is a monetary incentive attribute. In the pilot study stage, would you recommend using [+] to indicate preference order? (Other attributes do not have a clear preference order.)

3. If I treat the monetary incentive attribute as a continuous variable and the time attribute as a dummy variable (e.g., 5, 10, 15 min), is this acceptable? Or would it be better to treat both as dummy variables?

4. If there are a total of 14 parameters, then S >=7. Does the "7" refer to the total number of choice sets being > 7, or does each version need to be > 7?

5. If the pilot study uses rows=12, block=2, while the formal study uses rows=24, block=4, is this feasible?

Thank you in advance for your patient response.

Nancy

[Michiel Bliemer]

1. You may need to add an alternative-specific constants. If alt1 and alt2 are generic (of the same label), then you would have:

U(alt1) = b1*X + ...
U(alt2) = b1*X + ...
U(none) = b_optout

or you can put the constant in alt1 and alt2:

U(alt1) = b0 + b1*X + ...
U(alt2) = b0 + b1*X + ...
U(none) = 0

To set the utility of an alternative to zero in Ngene, you simply omit it in the utility specification (but still keep it in the ;alts specification).

If alt1 and alt2 are labelled alternatives, then you would have:

U(alt1) = b_alt1 + b1*X + ...
U(alt1) = b_alt2 + b1*X + ...
U(none) = 0

or

U(alt1) = b1*X + ...
U(alt1) = b_alt2 + b1*X + ...
U(none) = b_optout

Adding alternative-specific constants changes the number of parameters in the model and therefore changes the efficiency and D-error.

2. If the other attributes do not have a clear preference order, then you should NOT use [+] only for the monetary attribute because it is not possible to check for dominance when only one attribute has a preference order. You can simply leave all priors zero.

3. If you use zero priors, I generally recommend to dummy code all attributes, even numerical ones, to get more variation in the data. As explained, the manual, with zero priors it becomes optimal to compare only outer levels and inner levels across alternatives, which is not desirable. During model estimation, you can use a single coefficient for time and cost to indicate a continuous linear effect, even though you considered dummy coding during the design phase.

4. It refers to the design size (;rows in Ngene), so the total number of choice sets. How you divide it into blocks (versions) is not relevant for this calculation.

5. Yes, that is perfectly fine.

Michiel