Page 1 of 1

D-efficient design (unlabelled alternatives)

PostPosted: Thu Aug 13, 2020 10:06 pm
by Peter_C
Dear Michiel!

I would like to create a D-efficient design for three unlabelled alternatives.
I used the following syntax in Ngene:

design
;alts = alt1, alt2, alt3
;rows = 16
;block = 2
;eff = (mnl,d)
;con
;model:
U(alt1) = a1[0.001] + b2[-0.001] * price[10,18,26,30] + b3[-0.001] * invcost[200,300,700,1000] + b4[0.001] * env[1,2,3,4] + b5.dummy[-0.001] * conv[0,1] /
U(alt2) = a2[0.001] + b2*price + b3*invcost + b4*env + b5*conv /
U(alt3) = b2*price + b3*invcost + b4*env + b5*conv
$

I got the following results:

D error 0.009803
A error 0.180855
B estimate 86.613607
S estimate 1511311.361419


Is it possible to got that very large S estimate result (It can be caused by my relatively small prior values?)?

Furthermore in my results I found two problematistic choice tasks (due to dominant alternatives).
What is the best way/practice what I can do with these choice tasks?

Thank you very much for you answer!

Best regards,
Peter

Re: D-efficient design (unlabelled alternatives)

PostPosted: Fri Aug 14, 2020 10:15 am
by Michiel Bliemer
Hi Peter,

1. S-estimates are only meaningful if your priors come from a pilot study. Your priors are chosen just for sign (I think) so they are not meaningful, and therefore the S-estimates have no meaning and should be ignored.

2. You can tell Ngene to check for dominance by using ;alts = alt1*, alt2*, alt3*. This will not work with the default swapping algorithm because dominance checks impose heavy constraints, so you will need to swap to the modified Federov algorithm. This algorithm does not guarantee attribute level balance, so it is best to impose some constraints on the appearance of each attribute level.

For example:

Code: Select all
design
;alts = alt1*, alt2*, alt3*
;rows = 16
;block = 2
;eff = (mnl,d)
;alg = mfederov(candidates = 1000)
;con
;model:
U(alt1) = a1[0.001]
        + b2[-0.001] * price[10,18,26,30](2-6,2-6,2-6,2-6)
        + b3[-0.001] * invcost[200,300,700,1000](2-6,2-6,2-6,2-6)
        + b4[0.001] * env[1,2,3,4](2-6,2-6,2-6,2-6)
        + b5.dummy[-0.001] * conv[0,1](6-10,6-10)
        /
U(alt2) = a2[0.001] + b2*price + b3*invcost + b4*env + b5*conv /
U(alt3) = b2*price + b3*invcost + b4*env + b5*conv
$


Michiel

Re: D-efficient design (unlabelled alternatives)

PostPosted: Sat Aug 15, 2020 3:20 am
by Peter_C
Thank you very much for your help!