d-efficient design

[admin]

The following is being posted for a user for whom there was a technical issue posting to the forum software:

Dear all

I am a novice in learning Ngene. Please forgive me if the question I raised may be a little simple, I am doing research on physicians' preference for generic drugs.

There are no DCE or preference studies. I have designed 7 attributes myself (difference in response rate, disease severity, patient opinion, difference in adverse drug incidence, difference in drug interactions and considerations, price, policy), and want to use D-efficient design to generate selection sets.

difference in response rate=0,5%,10%,20%
disease severity=low,medium,high
patient opinion=support,oppose
difference in adverse drug incidence=0,0.1%,2%,5%
difference in drug interactions and considerations=no difference,different
price=no discount,discount 30%,discount greater than 50%
policy=no influence,influence

I wrote the syntax as follows:

Code: Select all

Design
;alts=drugA,drugB
;rows=24
;block=12
;eff=(mnl,d)
;model:
U(drugA)=effic*effic[0,0.05,0.1,0.2] + sever.dummy[0|0]*sever[0,1,2] + opinion.dummy*opinion[0,1] + inter.dummy*inter[0,1] + incid*incid[0,0.001,0.02,0.05] + policy.dummy*policy[0,1] + price.dummy[0|0]*price[0,1,2] /
U(drugB)=effic*effic + sever.dummy*sever + opinion.dummy*opinion + inter.dummy*inter + incid*incid + policy.dummy*policy + price.dummy*price
$

Could you please take a look at it for me and give me some advice?

I also want to ask a question about prior parameters. I saw a sentence in software manual: efficient designs rely on the accuracy of the prior parameter estimates. Must there be a prior parameter for D-efficient design?

Looking forward to your answer and thank you very much.

[Andrew Collins]

Overall the syntax looks reasonable (I made a few syntax corrections as admin).

I would definitely use ;alts=drugA*,drugB* to check for repeated choice tasks since you are using unlabelled alternatives.

The blocking means that each respondent only receives two choice tasks but perhaps that is what you want. If you want each respondent to see 12 choice tasks, then use ;rows=24 and ;block=2.

Prior parameters are not essential. If you have no prior information then you could use zero priors. If you run a pilot study, then you could estimate a model on the pilot data and use the parameter estimates for the main wave of the study.

If you can anticipate the sign (i.e. whether more of something generates positive or negative utility) then you might want to use values very close to zero that are positive or negative. Together with specifying ;alts=drugA*,drugB* this will also prevent dominance in the choice tasks, where one alternative is always better than or equal in performance to the other alternative.

Andrew

[wangqge]

Dear Andrew,

Thanks a lot for your answer.

I modified the syntax to run according to your suggestion.But I found that it would run for a long time and never stopped.What is going on?

Most of my attributes are qualitative. How can i judge whether it is positive utility or negative utility?Such as the opinion of a patient, it is positive if he supports generic drug ;it is negative if he is against it.How should i understand this situation.

I could determine the attribute 1 and attribute 6 have positive sign and attribute 2 and attribute 4 have negative signs.I am not sure about the others.

If I want to run a pilot study,can I conduct a pilot study based on the results of the current syntax design to obtain a prior parameters.

[Michiel Bliemer]

Effic: 4 levels
Sever: 3 levels
Opinion: 2 levels
Inter: 2 levels
Incid: 4 levels
Policy: 2 levels
Price: 3 levels

This means that drugA and drugB each have 4*3*2*2*4*2*3 = 1152 profiles. This means that there are 1152*1152 = 1,327,104 possible choice tasks.
You are asking Ngene to select the 24 best choice tasks out of 1,327,104 possible choice tasks. There are 1,327,104 * 1,327,103 * 1,327,102 * ... * 1,327,81 = 8,9049*10^146 possible combinations to do this. Therefore, Ngene will never stop trying to find a better design because it is impossible to get through this many designs unless you wait for several years. Therefore, you will need to stop the algorithm manually once you are happy with the design Ngene has found (I refer to the Ngene manual). For a relatively simple design like this, you can often stop it after a few minutes because Ngene will not be able to produce a much more efficient designs. Just keep an eye on the D-errors that Ngene reports while searching and you will see that after a few minutes only marginal improvements can be made.

Regarding positive/negative utilities, you typically cannot do this for nominal attributes (like "generic" or "colour") since there is no natural ordering and therefore you can set these priors at zero, but for example "price" has a clear attribute level order. If you can only determine the preference order of 2 attributes then there is no point in setting small positive or negative priors since dominant alternatives will not be an issue, so you can set them to zero for the pilot study. After the pilot study (typically using 10% of the total sample your budget can afford) you estimate the parameters and use these are priors to generate a new (more efficient) design.

I am not sure what the objective is of your study, but most people are interested in determining willingness to pay (WTP) for certain attributes and levels. Since you have no price attribute in dollars or another currency, you will not be able to compute WTP so you may want to change "discount" to an actual price level.

Michiel

[wangqge]

Dear Michiel,

Thank you very much for your patient answer.

I have a general understanding of software runs and pilot study.

The purpose of my study was to explore the factors the influence physicians' generic drug substitution and the extent to which these factors influence.

The two drug are generic drugs and brand-name drugs. Generic drugs and brand-name drugs are not designated as a class of drugs,so there can be a big difference in price.Generic drugs can range from a few dollars to thousands. So I use discounts to represent price differences.

[Michiel Bliemer]

My main issue with "discounts" is that people need to know the basis of the discount.

For example, suppose you are asked to choose between:

drugA: efficacy = 80%, discount = 10%
drugb: efficacy = 50%, discount = 50%

The respondent will not be able to make an informed choice here because the basis for the discount is unknown. Discount on a drug that costs around $1000 is very different from discount on a drug that costs $10. Of course it is possible to use scenarios to describe the context, for example:

Consider the drug for purpuse XXX with a base cost of $50.
drugA: efficacy = 80%, discount = 10%
drugb: efficacy = 50%, discount = 50%

Now discount is meaningful. However, in this case, you may just as well provide drug prices in dollars instead of asking respondents to calculate the discounts themselves.

Michiel

[wangqge]

Dear Michiel,

Thank you very much for the response,this will help me.