choice-metrics.com

Hello Professor, I am a graduate student.

Background: Two alternatives and a exit option. Six attributes three levels, where one attribute is a rating variable and there is a payment fee attribute.

Question 1: Which is better for wtp design or mlogit
; eff = (mnl, wtp(ref1))
; wtp = ref1(*/cost)
; model:
or
; eff = (mnl,d)
; model:

Question 2.1: Regarding the exit option, it seems that everyone will add ASC variable when browsing past posts. I imitate the code as follows (the prior value is the empirical hypothesis).

Design
; alts = choiceA*, choiceB*, neither
; rows = 24
; block = 2
; alg = mfederov
; eff = (mnl,d)
; model:

U(choiceA) = asc[0]
+cost[-0.001] * cost[3000, 300000,50000]
+ pre[0.001] * pre[0.2,0.4,0.6]
+ live[0.001] * live[0.15,0.35,0.55]
+ prisk[-0.001] * prisk[0.001,0.025,0.05]
+ brisk[-0.001] * brisk[0.05,0.1,0.15]
+ dec.dummy[0.002|0.001] * dec [1,2,0] ? 0 = low, 1 = mid, 2 = high
/
U(choiceB) = asc
+cost * cost
+ pre * pre
+ live * live
+ prisk * prisk
+ brisk * brisk
+ dec * dec
$

The result D error is Undefined. Is it a code error or an inaccurate prior value? (I try zero prior, D error=0.09)
P.S. Browsing past posts, I noticed that you would add (10-14,10-14,10-14) to the cost line. What does this do?

Question 2.2: My study is more suitable for "maintaining the status quo", that is, the subjects will not give up treatment, and at most express neutrality or indifference to the choice. Is there anything to pay attention to during the design phase and the data analysis phase?

Question 3: Acquisition of prior values.
I ultimately want to do efficient design, is the prior derived from the logit model of the orthogonal design? I saw A literature where the prior values were the Coefficients[Robust standard error] of model A in the Estimated mixed rank-ordered logit models.

I hope to get your reply. As a beginner, my questions may be some simple. Thanks again from my heart.

Not many people optimise for WTP, so using ;eff = (mnl,d) would be most common, but if you want to specifically optimise for estimating WTP then you could add this.

Please check the preference order of dec:

+ dec.dummy[0.002|0.001] * dec [1,2,0] ? 0 = low, 1 = mid, 2 = high

The last level (0) is the reference level hence 0 utility, but with your priors the utility flow mid (0.002) is larger than that of high (0.001).

You correctly added the constant.

The D-error is Undefined because your prior for cost is inappropriate; the cost levels are extremely large, therefore your prior must be very small as otherwise it completely dominates choice. Use -0.0000001 or something.

When using the modified Federov algorithm it is a good idea to impose constraints on how often each attribute level appears as otherwise the design may attribute level unbalanced. Since you have 3 levels, perfect attribute level balance would be if each level appears exactly 8 times across the 24 choice tasks. To achieve perfect attribute level balance, you could try adding (8,8,8) after the attribute levels. However, this would make it very difficult to find a feasible design, therefore it is a good idea to give upper and lower bounds, such as (6-10,6-10,6-10). This is mainly of importance for your numerical attributes, but is generally not needed for dummy coded categorical attributes since the D-error would be very low if certain levels do not appear in the design and therefore minimising the D-error automatically guarantees a high degree of attribute level balance for categorical variables.

I do not understand your question about status quo. You indicate that you have an opt-out alternative (Neither), but then you mention status quo. Note that a status quo alternative is not the same as an opt-out alternative. A status quo alternative has actual levels, whereas an opt-out alternative does not. For example in health you could have "surgery", "chemo" and "active surveillance" where the latter represents the no treatment option, but it still may have levels, e.g. no side effects and low life expectancy.

Priors typically come from a pilot study, but can be taken from the literature if you scale them properly. You could use an orthogonal design for your pilot study, but you can also use an efficient design with (near)zero priors as you seem to do. The latter has the advantage that it can automatically avoid dominant alternatives whereas an orthogonal design can not.

Michiel

Thank you very much for your answer, my project has finally made progress. Now I have consulted some clinical experts and come across a new problem.
My research is in the medical field. For example, if you spend 30,000 yuan, the cure rate is 30%, but if you spend 3,000 yuan, the cure rate is 60%.Is it unreasonable that you should have to choose between these two?
Can I modify it with some constraints? Or this is a normal part of DCE that need no correction?
Finally, I would like to express my heartfelt thanks to you.

Apologies for the late response.

You would typically apply constraints if you believe that certain attribute level combinations are unrealistic, but you always need to ensure that there is variation in the data as otherwise you cannot estimate the model. So you cannot link 30,000 yuan directly with a specific cure rate, but you could say that it belongs to several higher or lower cure rates.

Michiel