Correlation issues?

[jgrant8]

Hi everyone,

I am conducting an unlabelled DCE with 5 attributes with two alternatives plus an opt-out (a current policy option). Two of the attributes have 3 levels, and three of the attributes have 2 levels. My syntax is as follows:

Design
;alts = alt1, alt2, optout
;rows = 12
;eff=(mnl,d)
;model:
U(alt1) = asca[0.0001] + b1[0.0001]*mal[17,20,23] + b2.dummy[0.0001]*sev_mal[1,2] + b3.dummy[0.0001]*side_effects[1,0] + b4.dummy[-0.0001]*contacts[1,0] + b5.dummy[-0.0001|-0.0001]*dosing[2,1,0] /
U(alt2) = ascb[0.0001] + b1*mal + b2*sev_mal + b3*side_effects + b4*contacts + b5*dosing
$

I am coming up with two issues that I do not understand (and apologies if these are obvious, it is my first time doing a DCE and using Ngene):

1. When I look at Pearson's correlation matrix, there is high correlation (-1 for the first four attributes, and -0.5 for the dosing attribute) between each attribute in alternative 1 vs alternative 2. Is this a problem? And if so, how can I solve this?

2. The choice sets I get out all look okay, except for the first attribute- 'mal'/malaria. This is a 3 level (17, 20, 23) attribute, and in the choices situations, level 17 is always being compared with level 23, and level 20 always appears in both alternatives (see below). I am not sure why this is happening and what I can do to change it. I do not have this issue with my other 3 level attribute ('dosing').

alt1.mal alt2.mal
1 23 17
2 17 23
3 20 20
4 23 17
5 20 20
6 20 20
7 20 20
8 17 23
9 17 23
10 23 17
11 23 17
12 17 23

I am very grateful for any help!

[Michiel Bliemer]

1. This is perfectly fine and is in fact "optimal", you would also get the same correlations when you would use ;orth = ood. What matter is correlations WITHIN each alternative. Correlations across alternatives in an unlabelled experiment are not relevant.

2. This is what is expected and typically happens because it is most efficient with zero priors to compare as much as possible the outer levels of an attribute as this provides more Fisher information and efficiency. However, I agree that this is often not desirable, and there are two things that you could do. One, when using zero priors, using ;orth = ood. But my preferred way to deal with this is the dummy code ALL attributes, even the numerical ones. So just dummy code attribute MAL for now. You can still estimate it as a linear/numerical effect later, but for the design I would recommend to dummy code it, especially when your priors are (near) zero.
You can find more information about this in the new Ngene manual, page 88: https://files.choice-metrics.com/NgeneManual.pdf

I notice that you did not indicate in your script that alt1 and alt2 are unlabelled, therefore Ngene does not perform any dominance checks or repetition checks. I recommend you use ;alts = alt1*, alt2*, optout

Michiel

[jgrant8]

Thank you very much for your response! I have dummy coded MAL now as you suggested. And thank you for catching my error that I did not specify in the snytax that alt1 and alt2 are unlabelled.

Best,
Jane