Partial choice set design in a labelled DCE

[pakhi]

Hi Prof Bliemer,

I am planning to conduct a labelled DCE to explore face-to-face and virtual neurodevelopmental follow-up care for children in Australia. After conducting a literature review and qualitative exercises to narrow down the labels, attributes, and levels, we finalised five types of follow-up care alternatives:
1. Video conferencing
2. Telephone consultation
3. Mobile application
4. Online portal
5. Face-to-face (reference alternative)

Given that five alternatives might be overwhelming for respondents, I am considering using the partial choice set design (PCSD) method. My plan is to have the ‘face-to-face’ alternative appear in all choice sets, with any two alternatives from 1 to 4 appearing randomly in the choice sets, resulting in a total of three alternatives per choice task.

I understand that this can be achieved using an external candidate set as described in the Ngene choice metrics manual, where I could force the ‘face-to-face’ alternative to always appear in the external candidate set Excel file. Additionally, I am interested in exploring the labelled-to-unlabelled experiment reformulation method as discussed in this paper (https://onlinelibrary.wiley.com/doi/full/10.1002/hec.4666). Could this method also allow the ‘face-to-face’ alternative to consistently appear in all choice tasks?

I would greatly appreciate any guidance you can provide.

Thank you,
Pakhi

[Michiel Bliemer]

Yes, both methods can be used to generate a design with partial choice sets whereby one alternative always appears. The external candidate set method is the most flexible method and give you a lot of control over the choice sets by creating them for example in a spreadsheet.

The labelled-to-unlabelled reformulation method is a bit more tricky and requires quite a few clever constraints in Ngene but avoids having to generate an external candidate set. You would simply include the following 3 alternatives, where 2 are unlabelled and one is labelled:

* A (which can be of label video conference, telephone consultation, mobile application, and online portal)
* B (ditto)
* Face-to-face

Alternatively, you could use 3 unlabelled alternatives, A, B, C, where you simply impose constraints that C always has label dummy "Face-to-face", for example via
;require: C.label = 5

If your attributes are mostly the same across all alternatives, including mostly the same levels, then the labelled-to-unlabelled method would work quite well. But if your attributes are quite different across alternatives and/or have very different levels, then it is likely easier to use an external candidate set.

Michiel

[pakhi]

Thank you, that is very helpful!

[pakhi]

Hi Prof. Bliemer,

Following up on my previous question, I decided to use the external candidate method for generating a PCSD. I found a useful thread via this forum regarding candidate set generator for PCSD: viewtopic.php?f=2&t=882&p=4098&hilit=candidate+set#p3243

However, I am unable to access the spreadsheet link: https://cloudstor.aarnet.edu.au/plus/s/yTPQWAZpi2LFOCj

I tried opening it using different browsers but that didn’t work. I think (as per my knowledge) the problem might be with CloudStor because it was recently decommissioned by AARNet making links to CloudStor data inaccessible. Is there any other way I can access that spreadsheet?

Kind regards,
Pakhi

[Michiel Bliemer]

I will try to share with you via Dropbox. Can you try if this link works?
https://tinyurl.com/mtrrnamw

Michiel

[pakhi]

It worked, thank you so much Prof. Bliemer!

[pakhi]

Hi Prof. Bliemer,

Could I please ask a question regarding the same DCE design discussed in this thread?

For context, I am using five labelled alternatives (video, telephone, app, online, f2f). These are my attributes and levels: admin (0, 1, 2), tech (0, 1, 2), duration (30, 60, 90, 120), and feedback (0, 1). I am treating them all as alternative-specific parameters in the pilot test.

I generated a candidate set for PCSD where face-to-face(f2f) appears consistently, and any two alternatives from the remaining four vary, resulting in three choice sets per person. My Ngene codes are pasted below. While everything seems to work, the ;require property is not functioning as expected. E.g., I am getting choice sets where 'app.admin_app' is 1 or 2, but I need it to be 0, and similarly with all four ;require commands. I tried using the ;reject property as well, but the same issue persisted. Could this issue be due to having too many restrictions from the candidate set, or have I specified it incorrectly?

Ngene codes:

Code: Select all

Design
;alts = video, telephone, app, online, f2f
;rows = 50
;block = 5
;eff = (mnl, d)
;alg = mfederov(candidates = pcsd2.csv)

;require:
app.admin_app = 0,
online.admin_online = 0,
f2f.tech_f2f = 0,
f2f.duration_f2f >= 90

;model:
U(video) = video[0]
+ b1_video.dummy[0.000001 | 0.000002] * admin_video[0, 1, 2]
+ b2_video.dummy[0.000001 | 0.000002] * tech_video[0, 1, 2]
+ b3_video[-0.000001]                 * duration_video[30, 60, 90, 120, 999]
+ b4_video.dummy[0.000001]            * feedback_video[0, 1]
/

U(telephone) = telephone[0]
+ b1_telephone.dummy[0.000001 | 0.000002] * admin_telephone[0, 1, 2]
+ b2_telephone.dummy[0.000001 | 0.000002] * tech_telephone[0, 1, 2]
+ b3_telephone[-0.000001]                 * duration_telephone[30, 60, 90, 120, 999]
+ b4_telephone.dummy[0.000001]            * feedback_telephone[0, 1]
/

U(app) = app[0]
+ b1_app.dummy[0.000001 | 0.000002] * admin_app[0, 1, 2]
+ b2_app.dummy[0.000001 | 0.000002] * tech_app[0, 1, 2]
+ b3_app[-0.000001]                 * duration_app[30, 60, 90, 120, 999]
+ b4_app.dummy[0.000001]            * feedback_app[0, 1]
/

U(online) = online[0]
+ b1_online.dummy[0.000001 | 0.000002] * admin_online[0, 1, 2]
+ b2_online.dummy[0.000001 | 0.000002] * tech_online[0, 1, 2]
+ b3_online[-0.000001]                 * duration_online[30, 60, 90, 120, 999]
+ b4_online.dummy[0.000001]            * feedback_online[0, 1]
/

U(f2f) =
  b1_f2f.dummy[0.000001 | 0.000002] * admin_f2f[0, 1, 2]
+ b2_f2f.dummy[0.000001 | 0.000002] * tech_f2f[0, 1, 2]
+ b3_f2f[-0.000001]                 * duration_f2f[30, 60, 90, 120, 999]
+ b4_f2f.dummy[0.000001]            * feedback_f2f[0, 1]
$


Thank you very much.

Kind regards,
Pakhi

[Michiel Bliemer]

When you are using an external candidate set, I am not sure if Ngene actually applies the constraints as the assumption would be that all choice sets in the candidate set are allowed. So I would recommend applying any other constraints directly in the candidate set in the spreadsheet and not in the script in Ngene.

Michiel

[pakhi]

Hi Prof. Bliemer,

Thank you for your advice. I was able to generate a design using a pcsd candidate set imposing all relevant constraints within it. However, the attribute level balance (image in this link: https://docs.google.com/document/d/1f228JnGFldtxYbg1cSoWuUy57bkL2d_u-h_A8-kfWKA/edit?usp=sharing), as expected, was not good, possibly due to the imposed constraints and usage of mfederov algorithm. Would you recommend manually changing the levels of some attributes to obtain a suitable level balance?

Thank you.

Kind regards,
Pakhi

[Michiel Bliemer]

You can try including attribute level balance constraints, for example:

attribute[1,2,3](10-25,10-25,10-25)

whereby each level must appear between 10 and 25 times in the design. To determine appropriate values, you first divide the number of rows (in your case 50) with the number of levels (in this case 3) to get a number for perfect balance. And then you set values around such perfect balance.

Note that the more restrictions you impose on your design, the more difficult it will be for Ngene to find a design that satisfies all your constraints. So if Ngene does not find a design, you need to relax your constraints by widening the ranges.

Michiel