Decision situations with dominan alternative

[Peter_C]

Hi Michiel,

I have tried to run two design in Ngene and I have some problem.

For design 1:

I have 4 attributes with 2-2-2-5 levels (in unlabelled case).

My design is the following:

design
;alts = Service_A, Service_B, Service_C, Service_D
;rows = 8
;block = 2
;eff = (mnl,d)
;model:
U(Service_A) = b2[0.0001] * int[0,1]
+ b3[0.0001] * iter[0,1]
+ b4[0.0001] * mobil[0,1]
+ b5[-0.0001] * cost[10,20,30,40,50]
/
U(Service_B) = b2*int + b3*iter + b4*mobil + b5*cost /
U(Service_C) = b2*int + b3*iter + b4*mobil + b5*cost /
U(Service_D) = b2*int + b3*iter + b4*mobil + b5*cost
$


Int, inter and mobil attributes have only two levels (0-no,1-yes), so I used linear instead of dummy coding.

I haven't got any informations about prior values. This research will be only a pilot study before the final research.

In this case, I get a lot of decision situations that contain dominant alternative.

I have tried to use "*"signal, but I get an error message (I have too much dominant alternatives).

What can I do in this case?

For design 2

I have 7 attributes with 2-2-2-2-2-2-5 levels (in unlabelled case).

My design is the following:

design
;alts = Service_A, Service_B, Service_C, Service_D
;rows = 8
;block = 2
;eff = (mnl,d)
;model:
U(Service_A) = b2[0.0001] * at[0,1]
+ b3[0.0001] * in[0,1]
+ b4[0.0001] * mo[0,1]
+ b5[0.0001] * bi[0,1]
+ b6[0.0001] * ie[0,1]
+ b7[0.0001] * we[0,1]
+ b8[-0.0001] * cost[10,20,30,40,50]
/
U(Service_B) = b2*at + b3*in + b4*mo + b5*bi + b6*ie + b7*we + b8*cost /
U(Service_C) = b2*at + b3*in + b4*mo + b5*bi + b6*ie + b7*we + b8*cost /
U(Service_D) = b2*at + b3*in + b4*mo + b5*bi + b6*ie + b7*we + b8*cost
$

This is a same situation, I get a lot of decision situations with dominant alternatives. Every attributes have two levels (0:no,1-yes), except the cost.

This is also a pilot study.

What can I do in this cases?

I can not increase the number of levels, because these are really a two level attributes.


Thank you very much for your help!

Peter

[Michiel Bliemer]

This can be fixed by changing the algorithm in Ngene, see syntax below where I added comments to highlight the changes.

Code: Select all

design
;alts = Service_A*, Service_B*, Service_C*, Service_D*  ? added asterisks
;rows = 8
;block = 2
;eff = (mnl,d)
;alg = mfederov   ? changed to Modified Federov algorithm, which is much better in handling severely constrained designs
;model:
U(Service_A) = b2[0.0001] * int[0,1]
+ b3[0.0001] * iter[0,1]
+ b4[0.0001] * mobil[0,1]
+ b5[-0.0001] * cost[10,20,30,40,50](1-2,1-2,1-2,1-2,1-2)  ? added constraints to ensure that each level appears 1-2 times since the mfederov algorithm cannot maintain attribute level balance
/
U(Service_B) = b2*int + b3*iter + b4*mobil + b5*cost /
U(Service_C) = b2*int + b3*iter + b4*mobil + b5*cost /
U(Service_D) = b2*int + b3*iter + b4*mobil + b5*cost
$

You can do something similar for your second design.

Michiel

[Peter_C]

Dear Michiel!

Thank you very much for your help!

It worked!

In meantime, another question arised in me. Another approach.

In this case, I have only 2 attributes (with 3-5 levels).

What is the difference if I handle the two-level attributes as the following:

design
;alts = Service_A, Service_B, Service_C, Service_D
;rows = 8
;block = 2
;eff = (mnl,d)
;model:
U(Service_A) = b2.dummy[0.0001|0,0001] * twolevelattributes[int,iter,mobil]
+ b3[-0.0001] * cost[10,20,30,40,50]
/
U(Service_B) = b2*twolevelattributes + b5*cost /
U(Service_C) = b2*twolevelattributes + b5*cost /
U(Service_D) = b2*twolevelattributes + b5*cost
$

In your opinion, which one is the better solution?

Thank you for your answer!

Best wishes,
Peter

[Michiel Bliemer]

With these 3 levels you cannot replicate all combinations in your three 2-level attributes, but perhaps I misunderstand what you are trying to do. Moving three attributes into a single one will require 6 levels, so you end up with the same number of parameters. Keeping them as separate dummy variables is the appropriate way of dealing I think.

Michiel

[Peter_C]

Maybe a silly question and probably I asked strangely.

I try to creat a maybe better understandable example below.

If I want to analyse three different type of labels (label of origin, fairtrade label, bio label).

What is the main difference between the following ways:

1. I separately use this labels in utility specification as different attributes (in this case I estimate parameters for origin, fairtrade, bio and cost (4 parameter) (base levels are: no origin, no fairtrade, no bio)):

design
;alts = Product_A, Product_B, Product_C
;rows = 8
;block = 2
;eff = (mnl,d)
;model:
U(Product_A) = b2[0.0001] * origin[0,1]
+ b3[0.0001] * fairtrade[0,1]
+ b4[0.0001] * bio[0,1]
+ b5[-0.0001] * cost[10,20,30,40,50]
/
U(Product_B) = b2*origin + b3*fairtrade + b4*bio + b5*cost /
U(Product_C) = b2*origin + b3*fairtrade + b4*bio + b5*cost
$

2. I creat and use a new attribute, for example label, which have three different type of levels (in this case I estimate parameters for origin, fairtrade and cost (3 parameter) (base level is bio):

design
;alts = Product_A, Product_B, Product_C
;rows = 8
;block = 2
;eff = (mnl,d)
;model:
U(Product_A) = b2.dummy[0.0001|0.0001] * label[origin,fairtrade,bio]
+ b3[-0.0001] * cost[10,20,30,40,50]
/
U(Product_B) = b2*label + b3*cost /
U(Product_C) = b2*label + b3*cost
$



So my 2-level attributes are very similarly. I can handle these separately and merge them into one attributes.

What is the main difference in terms of these two approach?

I don't know if my question makes sense.

Peter

[Michiel Bliemer]

The first syntax allows you to analyse 8 combinations of characteristics by allowing each attribute to be true or false:

1. Origin = Yes, Fairtrade = Yes, Bio = Yes
2. Origin = Yes, Fairtrade = Yes, Bio = No
3. Origin = Yes, Fairtrade = No, Bio = Yes
4. Origin = Yes, Fairtrade = No, Bio = No
5. Origin = No, Fairtrade = Yes, Bio = Yes
6. Origin = No, Fairtrade = Yes, Bio = No
7. Origin = No, Fairtrade = No, Bio = Yes
8. Origin = No, Fairtrade = No, Bio = No

In the second syntax you only allow origin OR fairtrade OR label, but not combinations of them, leading to only the following 3 combinations of characteristics:

1. Origin = Yes, Fairtrade = No, Bio = No
2. Origin = No, Fairtrade = Yes, Bio = No
3. Origin = No, Fairtrade = No, Bio = Yes.

Further, in the second syntax, the reference is Bio, so the estimated coefficients describe the importance of origin and fairtrade relative to Bio, whereas in the first syntax you estimate coefficients that express separate contributions to utility of each characteristic.

Michiel

[Peter_C]

Thank you very much!

It became absolutely clear.

Best regards,
Peter

[Peter_C]

Dear Michiel!

I would have another question about the Modified Federov algorithm.

How can we determine the values of constraints in case of Modified Federov algorithm.

Is it based on the number of alternatives in decision situations, and the number of attribute levels?

You suggested above that to add the following constraint for the cost attribute: (1-2,1-2,1-2,1-2,1-2)

What is based on?

For example, what values should I enter for a three-level attribute if I have 16 decision situations?

Thank you very much for your answer.

Peter

[Michiel Bliemer]

If S is the number of rows, and L is the number of levels of an attribute, then you would have perfect attribute level balance if each level appears S/L times. Therefore, you need to select a range around S/L.

In case of 16 choice tasks and 3 levels, 16/3 = 5.33, so you can use 5-6 as a range but that is quite restrictive and Ngene may not be able to find a design, so then you would increase the range to 4-6 or 4-7. The larger the range, the more likely that the modified Federov algorithm will be able to generate a feasible design.

Michiel

[Peter_C]

Became clear!

Thank you very much!