choice-metrics.com

Hi,
Currently I am working on research project entitled “U. S. Consumers’ willingness to pay for pork produced with minimal use of antibiotics or no use of antibiotics.” I am conducting discrete choice experiment and want to use efficient designs to find out the choice sets. In our research we have 3 alternatives; Option A, Option B and Neither. The four attributes are Price ($2.99, $4.99, $6.99 and $8.99), Use of Antibiotics (Conventional, Minimal Use and Antibiotic free), Production Method (Standard, Traditional) and Use of synthetic growth promoter (SGP) (Yes, No) with their attribute levels within the parentheses.

We have developed the model like this to run in Ngene,

Design
;alts= optA, optB, Neither
;rows=8
;eff= (rp,d, mean)
;block=2
;model:
U(optA) = b1[(u,-0.34, -0.28)]*price.ref[2, 4, 6, 8]
+ b2[u, -0.8, 1.1] * Anti.ref[0, 2, 4]
+ b3[(u, 0.5,1)] * production.ref[1, 2]
+ b4[(u,-0.7, 0.5)] * SGP.ref[0, 1]/

U(optB) = b1*price[2, 4, 6, 8]
+ b2 * Anti[0, 2, 4]
+ b3 * production[1, 2]
+ b4 * SGP[0, 1] /

U(Neither) = b1*price[0]
+ b2 * Anti[0]
+ b3 * production[0]
+ b4 * SGP[0]
$

I am really not sure about the model we designed to run in Ngene. I am very new in using the Ngene software. Could you let me know whether I am going on right track or do I need to make the modifications on it??

A few comments:

1. Utility for 'Neither' should likely have a constant and no other attributes
2. You will likely want to optimise for the MNL model, not RP
3. Some variables require dummy coding
4. Think about your priors
5. Do you need the 'neither' option? You do not need it for willingness-to-pay studies.

May use the syntax below as a starting point:

Code: Select all

Design
;alts= optA, optB, Neither
;rows=16
;eff=(mnl,d)
;block=2
;model:
U(optA) = b1 * price[2, 4, 6, 8]
        + b2.dummy[0|0] * Anti[2,1,0]
        + b3.dummy[0] * production[1,0]
        + b4.dummy[0] * SGP[1,0] /
U(optB) = b1 * price
        + b2 * Anti
        + b3 * production
        + b4 * SGP /
U(Neither) = b0
$

Michiel

Hello,
Thank you so much for your response. It is really very helpful for me. Still I have some of queries related to the design.

1. If we use multinomial logit (mnl) to optimize the choice sets, can we use Random parameter model during data analysis? (We want to analyze our survey data using random parameter model later)
2. What is the reason of having 16 rows instead of 8 rows?
3. How do we include prior values when we have dummies for attribute levels?
4. Is there any provision in Ngene that we can export the choice sets from Ngene to survey platform such as QuestionPro?

Your suggestions regarding this matter will be highly appreciated.

Hoping to hearing from you soon.

1. In most cases the answer is "yes". To be sure, you can evaluate the MNL-optimised design for an RPPANEL model in Ngene. Note that optimising for the RPPANEL model is extremely computationally intensive and in most cases you are better off optimising for the MNL model, noting that an MNL efficient design is generally also RPPANEL efficient.

2. I think that a design with 8 rows is quite small for estimating 6 parameters in your model. I think that you may want to use a larger design in order to create more variation in your data. But if you prefer 8 rows then that is fine.

3. Please see my syntax where I added zero priors for the dummy coded coefficients. But maybe I misunderstand your question?

4. No. All survey instruments are different with respect to the format of the design, which means that you will likely need to reformat the design yourself. This is typically easy to do in Excel. Some survey instruments can read Ngene designs directly, e.g. SurveyEngine.

Michiel

Thank you so much for your time. It is really very helpful to me.

bpaudel

Hi,
As per your suggestions, i revised the model as below. Can you please suggest me if there any problem in the model?

Design
;alts= optA, optB, Neither
;rows=16
;eff=(mnl,d, mean)
;block=2
;model:
U(optA) = b1 [(n,-0.4, 0.03)] * price[2, 4, 6, 8]
+ b2.dummy[(n, 1.3, .70)| (n, 0.4, 0.04)] * Anti[2, 1, 0]
+ b3.dummy[(n, 0.8, .07)] * production[1,0]
+ b4.dummy[(n, 0.4, 0.01)] * SGP[1,0] /
U(optB) = b1 * price
+ b2 * Anti
+ b3 * production
+ b4 * SGP /
U(Neither) = b0
$

Further, please let me know when we run the model, current evaluation goes on and on. Do i need to stop the action at some point or i need to wait until it stops? Also, i am getting different choice sets when running the same model. I am confused. Please, give me an insight related to these matters.

With regards,
bpaudel

The constant for the no-choice has an important impact on the choice probabilities and hence the efficiency of the design. It is currently set to zero, but you may want to update based on estimated from a pilot study.

You may also want to avoid dominant alternatives by setting ;alts = optA*, optB*, Neither

Ngene will continuously look for a better design, but because there are trillions of different possible designs it will just keep evaluating more designs until you stop the algorithm. Any design you select is fine for use (and they will all be different), but the lower the D-error the more efficient the design. If the D-error does not go down much anymore then you can simply stop the algorithm. This may happen already after a few minutes.

Michiel

Hi Michael,
Thank you so much for your suggestions.

Is there any provision for improving our priors after pilot study in Ngene? Please kindly let me know if Ngene does so.

With regards,
bpaudel

I am not quite sure I understand your question. Once you have obtained priors for the coefficients after a pilot study, you simply generate a new design using these new (Bayesian) priors in the same way as you have entered your (Bayesian) priors in the syntax now. You can use the parameter estimate as the mean and the standard error as the standard deviation in a normally distributed Bayesian priors.

Michiel

Hi Michael,
Thank you so much for your response. I can relate your point to my study.

alts = optA*, optB*, Neither

Can you please tell me, how is this different from what you have without star? Does this create any difference in results?

Your suggestions regarding this matter will be highly appreciated. Hoping to hearing from you soon.

With regards,
bpaudel