Neither choice is disappeared on a formatted scenario

[Poonyawee]

Hi
I have some question about program. I will describe a little bit about my data which consist of 5 attributes
1. Efficacy (low, moderate, high)
2. GI side effect (Yes, No)
3. Kidney side effect (Yes, No)
4. Slow disease progression (None, Moderate)
5. Price per month (<500, 501-1000, >1000)
I would like to generate 2 alternative and neither.
So, this is my syntax

Design
? This will generate a simultaneous orthogonal fractional factorial design with three blocks Design
;alts = Drug A, Drug B, Neither
;rows = 36
;orth = sim
;block = 6
;model:
U(Drug A) = b1 + b2 * Efficacy[1,2,3] + b3 * GI side effect[1,2] + b4 * Kidney side effect[1,2] + b5 * Slow disease progression[1,2] + b6 * Price per month[500,501,1000]/
U(Drug B) = b2 * Efficacy[1,2,3] + b3 * GI side effect[1,2] + b4 * Kidney side effect[1,2] + b5 * Slow disease progression[1,2] + b6 * Price per month[500,501,1000]/
U(Neither) = b0
$

My questions are....
1. On a formatted scenario, it is only show 2 alt but neither choice is disappeared. What should I do?
2. How can I define an appropriate row and block?
3. How about sample size calculation? Depend on this syntax, sample size should be at least 60 people?
4. Should I put ;foldover into this syntax?

Thank you

[Michiel Bliemer]

1. You can add rows or columns by right-clicking on the choice scenario layout table. The neither alternative does not have attribute so you can leave the rest of the table empty. You can format the new row or column by selecting "cell style" below the table.

2. What do you mean with "an appropriate row and block"?

3. Sample sizes are case specific and depends on both the attribute levels as well as appropriate parameter priors (e.g. coming from a pilot study). Without priors it is not possible to make an estimate of the required sample size.

4. No need to use foldover since you do not have interaction terms.

Note that you cannot use spaces in the variable names so place remove all spaces within "GI side effects" etc.
Note that you need to use dummy or effects coding for qualitative variables, e.g. b3.dummy[0] * sideeffects[1,2]. where 2 is the reference level.
Note that DrugA and DrugB seem unlabelled alternatives, therefore it should not have a constant b1 for DrugA.
Note that if you want to check for dominance you can use small positive or negative priors (e.g. b2[0.00001] * efficacy[1,2,3]) and use ;alts = DrugA*, DrugB*.
Note that for an unlabelled experiment it is more appropriate to use a sequential orthogonal design than a simultaneous orthogonal design.
Note that instead of using an orthogonal design you could also opt for an efficient design with zero priors by using ;eff = (mnl,d).

Michiel

[Poonyawee]

Dear Ngene expert
Thank you very much for your suggestion in previous reply.
This is my revised syntax:
Design
? This will generate a simultaneous orthogonal fractional factorial design with three blocks Design
;alts = Drug A, Drug B, neither
;rows = 36
;eff = (mnl,d).
;orth = seq
;block = 6
;model:
U(Drug A) = b2 * Efficacy[0,1,2] + b3.dummy[0] * GISideEffects[0,1] + b4 .dummy[0] * KidneySideEffect[0,1] + b5.dummy[0] * SlowDiseaseProgression[0,1] + b6 * PricePerMonth[500,750,1000]/
U(Drug B) = b2 * Efficacy[0,1,2] + b3.dummy[0] * GISideEffects[0,1] + b4 .dummy[0] * KidneySideEffect[0,1] + b5.dummy[0] * SlowDiseaseProgression[0,1] + b6 * PricePerMonth[500,750,1000]/
U(neither) = b0
$

I have some question about my attribute which is a symbol --> Price per month (<500, 501-1000, >1000)
In my understanding, we cannot write these symbol (<,-,>) into syntax. So could I change to assume number such as--> b6 * PricePerMonth[500,750,1000]? Then, I will revert to the old one on the formatted scenario. Is it any effect to design?

I am a beginner. Please give me advice.

Thanks in advance for any help that you are able to provide.

[Michiel Bliemer]

You are proposing to use price as a categorical variable ("<500", "501-1000", ">1000") instead of a continuous variable. I would not recommend this approach since the issue is exactly as you indicate, it is unclear which value the respondent has in mind when you show these ranges. Instead, it is more appropriate to use specific price levels that you show to respondents to rule out any vagueness about the price level. If you insist on including price using these categories, you will need to use dummy or effects coding, e.g. b6.dummy or b6.effects.

Michiel

[Poonyawee]

Dear Michiel and Ngene expert

Thank you very much for your suggestion in previous reply. Therefore, I decide to use continuous variable as your recommendation.
I try to think about my syntax in case of no data prior. And My additional question is that
1. Efficacy attribute is 3 level (low moderate high). Should I use dummy coding?
2. If I insist to use zero to form DCE for all respondent, it is possible or any impact to design?
3. I would like to confirm syntax with you whether true or not.

This is my revised syntax:
Design
;alts = Drug A*, Drug B*, neither
;rows = 36
;eff = (mnl,d) or ;orth = seq
;block = 6
;model:
U(Drug A) = b1 * Efficacy[0,1,2] + b2.dummy[0] * GISideEffects[0,1] + b3 .dummy[0] * KidneySideEffect[0,1] + b4.dummy[0] * SlowDiseaseProgression[0,1] + b5 * PricePerMonth[500,750,1000]/
U(Drug B) = b1 * Efficacy[0,1,2] + b2.dummy[0] * GISideEffects[0,1] + b3 .dummy[0] * KidneySideEffect[0,1] + b4.dummy[0] * SlowDiseaseProgression[0,1] + b5 * PricePerMonth[500,750,1000]/
U(neither) = b0
$

4.If I have a data from pilot study, I should use these number and syntax will be like this --> b1[....] * Efficacy[0,1,2] + b2.dummy[....] * GISideEffects[0,1]
Is it true?


Thanks in advance for any help that you are able to provide.

[Michiel Bliemer]

1. I would suggest using dummy coding, yes.
2. I do not understand this question, could you please reformulate?
3. See my suggestion for syntax below.
4. You will need to replace the prior values for the parameters (e.g. -0.0001) with the parameter values obtained in model estimation based on your pilot data. If B is the parameter value and SE is the standard error for the parameter, then you either use b[B] for local priors, or b[(n,B,SE)] for Bayesian priors. The latter approach is preferred in order to make the design more robust against prior mis-specification.

Proposed syntax:

Code: Select all

Design
;alts = DrugA*, DrugB*, neither
;rows = 36
;eff = (mnl,d)
;alg = mfederov
;block = 6
;model:
U(DrugA)   = b1.dummy[0.0001|0.0002] * Efficacy[1,2,0]                                 ? 0 = low, 1 = medium, 2 = high
           + b2.dummy[-0.0001]       * GISideEffects[1,0]                              ? 0 = no, 1 = yes
           + b3.dummy[-0.0001]       * KidneySideEffect[1,0]                           ? 0 = no, 1 = yes
           + b4.dummy[0.0001]        * SlowDiseaseProgression[1,0]                     ? 0 = no, 1 = yes
           + b5[-0.0001]             * PricePerMonth[500,750,1000](10-14,10-14,10-14)  ? requiring that each level appears between 10 and 12 times
           /
U(DrugB)   = b1 * Efficacy
           + b2 * GISideEffects
           + b3 * KidneySideEffect
           + b4 * SlowDiseaseProgression
           + b5 * PricePerMonth
           /
U(neither) = b0[0]
$

Michiel

[Poonyawee]

Dear Michiel and Ngene expert

Thank you very much for fast response and sorry for unclear question.

1. It means that if I do not do a pilot study, It is possible to use syntax as your recommendation for total sample size. Is there any effect to the design.

2. Please explain me about this syntax --> + b5[-0.0001] * PricePerMonth[500,750,1000](10-14,10-14,10-14) ? requiring that each level appears between 10 and 12 times
Why we should use negative prior instead of positive prior?
What dose it mean of "10-14' and "requiring that each level appears between 10 and 12 time".

3. I quite understand how to write syntax with positive prior from your example but I wonder about negative prior, if level 0 is better than level 1, and level 1 is better than level 2 and level 2 is better than level 3. How to write syntax? -->
+ b10.dummy[-0.0001|-0.0002|-0.0003] * B[1,2,3,0] or .
+ b10.dummy[-0.0003|-0.0002|-0.0001] * B[3,2,1,0] or
+ b10.dummy[-0.0003|-0.0002|-0.0001] * B[3,2,1,0] or
+ b10.dummy[0.0003|0.0002|0.0001] * B[0,1,2,3] or...

Could you please explain how to write negative prior with dummy coding?

Sorry for a lot of questions and thanks in advance for any help that you are able to provide.

[Michiel Bliemer]

1. You can use the syntax I provided if you do not do a pilot study. However, without appropriate priors from a pilot study the sample size estimates provided by Ngene should be ignored. It is not possible to provide a sample size estimate without appropriate priors.

2. Sorry I meant to say between 10 and 14 times. I refer to page 146 of the Ngene manual, as well as page 221 that explains the attribute level constraints. Price represents a disutility, the higher price, the lower utility, therefore there is a negative relationship between price and utility, such that the parameter is expected to have a negative sign.

3. There are multiple ways of doing this, it depends on which level you choose as reference level and in which order you define the levels. All these specifications do the same thing:
b.dummy[0.0003|0.0002|0.0001] * B[0,1,2,3], which uses level 3 as the reference level
b.dummy[-0.0001|-0.0002|-0.0003] * B[1,2,3,0], which uses level 0 as the reference level
b.dummy[-0.0003|-0.0002|-0.0001] * B[3,2,1,0], which uses level 0 as the reference level

Michiel