Ngene Syntax

[ENich]

Hi,

I'm carrying out a DCE to determine the factors that influence GP decision-making when referring children to the emergency department. Given the nature of the study (many of the attributes are patient attributes), we will not be asking GPs to trade two patients off one another but will design a DCE that will present a single presentation profile (i.e., no alternative set of choices) and they will be provided with a binary response option of "Refer" or "Not Refer". I have never designed a DCE like this but I have seen them used in the literature.

We are employing a D-efficient design with 12 choice sets over 2 blocks. There is very little literature in this area and no other DCEs (as far I as can see), therefore, we will have to wait until we have collected pilot data before entering the priors.

I have put together the syntax below which is running in Ngene for me and appears to generate the choice sets as needed for the study, however, I was wondering if you could take a look and see if there is anything I am missing or have overlooked. Let me know if you need any further details about the study.

Ngene syntax:

Design
;alts = alt1, alt2
;rows = 12
;block = 2
;eff = (mnl,d)
;model:
U(alt1) = b0[0] + b1.dummy[0|0]*TimeofReferral[1,2,0] + b2.dummy[0]*PaediatricClinic[1,0] + b3.dummy[0|0]*ReferralRequest [1,2,0] + b4.dummy[0|0]*Capacity[1,2,0] + b5.dummy[0]*Competence[1,0] + b6.dummy[0]*Repeat[1,0]
$

Thanks,

E

[Michiel Bliemer]

That syntax looks good. A few comments:

1. Your experiment has two labelled choice alternatives without any profiles, but rather has a patient scenario that changes. The patient is the scenario, while the choice alternatives do not have any attributes. Can you confirm that all variables are patient characteristics, i.e. that they are scenario variables rather than attributes of the alternatives? This looks similar to a study that I was involved in here: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0142507

2. I made your syntax look a bit better by choosing appropriate variable names and I also suggest that you always use comments to indicate what each dummy coded level means

3. You are estimating 10 parameters, then using 12 rows in your design is not a lot. I suggest you increase the variation in your data set by using 24 rows and block it in 4 versions. Note that these choice tasks are relatively simple so a GP may be able to answer more than 6.

4. I added ;con because I think you will want to optimise on estimating the constant as well.

Code: Select all

Design
;alts = refer, not_refer
;rows = 24
;block = 4
;eff = (mnl,d)
;con
;model:
U(refer) = refer_asc[0]
         + tor.dummy[0|0] * TimeofReferral[1,2,0]  ? 0 = ... (ref) 1 = ... 2 = ...
         + pc.dummy[0]    * PaediatricClinic[1,0]  ? 0 = ... (ref) 1 = ...
         + rr.dummy[0|0]  * ReferralRequest[1,2,0] ? 0 = ... (ref) 1 = ... 2 = ...
         + cap.dummy[0|0] * Capacity[1,2,0]        ? 0 = ... (ref) 1 = ... 2 = ...
         + comp.dummy[0]  * Competence[1,0]        ? 0 = ... (ref) 1 = ...
         + rep.dummy[0]   * Repeat[1,0]            ? 0 = ... (ref) 1 = ...
$

Michiel

[ENich]

Hi Michiel,

Thank you so much this is really helpful. Just a follow up on a few comments below:

1. Yes I can confirm that all attributes are scenario variables. There will be fixed variables (age of child, symptom profile etc.) that will provide context to the scenario but are not attributes.

3. The plan is to have each GP answer 12 choice sets which is why I included two blocks but, apologies, this was an error on my part. I had included two blocks because a second component of the study will examine the same decision making but the child will have a disability (which will be a fixed variable) and I had planned on assigning the one block to include the scenarios where the child has a disability and the other where they do not (each GP will receive 6 of each). If possible, we aim to carry out two separate analyses on these two blocks rather than include disability as an attribute. Do I need to build this into the syntax?

Thanks again,

E

[Michiel Bliemer]

Regarding adding disability, since you are using zero priors and you intend to estimate two separate models, there is no need to include disability in the syntax.

I would still suggest using 24 rows or more in order to create sufficient variation in your data, especially if you are estimating separate models.
Suppose you create a design with 24 rows and 4 blocks, say B1, B2, B3, B4, each having 6 choice tasks (without disabilty).

If you give GP1 two blocks, namely B1(no disability) and B2(disability), and GP2 two other blocks, namely B3(no disability) and B4(disability), then each data set will only have 2 blocks (12 different choice tasks) for estimating a model. This requires creating 2 versions of the survey.

If you give GP1 blocks B1(no disability) and B1(disability), GP2 blocks B2(no disabilty) and B2(disability), GP3 blocks B3(no disability) and B3(disability), and GP4 blocks B4(no disability) and B4(disability), then each data set will have 4 blocks (24 different choice tasks) for estimating a model. This requires creating 4 versions of the survey. By having the same blocks with and without disability you also have more statistical power if you want to test whether disability has a significant impact on choice.

Michiel