Paired Choice Fractional Experiment Design Question

[Perdue, Richard]

Hi Folks

 

My apologies if you've talked about this in the past.  I'm a newcomer to
this list.  

 

I am doing a project examining consumer preferences and would like to
use a fractional factorial design to do a paired choice preference
study.

I have five attributes with (4)(2)(4)(2)(4) levels.  I am able to use
the SPSS design function to create an orthogonal, main effects only
model for one presentation of these attributes, but am uncertain how to
do the paired presentation.  

 

Your advice would be greatly appreciated.

 

Thanks

 

Rick Perdue

Virginia Tech

[Anthony Babinec]

ORTHOPLAN was designed with rank- or ratings-based conjoint
analysis in mind, and not choice modeling. However, you should
be able to make ORTHOPLAN work for you by doubling your number
of columns - that is, generate an orthogonal design that has columns

att1_alt1 att2_alt1 ... att5_alt1 att1_alt2 att2_alt2 ...att5_alt2

For more on the use of ORTHOPLAN in choice models, see "Applied
Choice Analysis" by David Hensher, John Rose, and William Greene,
published by Cambridge University Press.

Anthony Babinec
[hidden email]

"Be the change you want to see in the world." Gandhi

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Perdue, Richard
Sent: Wednesday, October 03, 2007 12:02 PM
To: [hidden email]
Subject: Paired Choice Fractional Experiment Design Question

Hi Folks



My apologies if you've talked about this in the past.  I'm a newcomer to
this list.



I am doing a project examining consumer preferences and would like to
use a fractional factorial design to do a paired choice preference
study.

I have five attributes with (4)(2)(4)(2)(4) levels.  I am able to use
the SPSS design function to create an orthogonal, main effects only
model for one presentation of these attributes, but am uncertain how to
do the paired presentation.



Your advice would be greatly appreciated.



Thanks



Rick Perdue

Virginia Tech

[Perdue, Richard]

Thanks Anthony

I have the Hensher, Rose and Greene book as well as the one by Louviere,
Hensher and Swait "Stated Choice Methods" and Train's "Discrete Choice
Methods with Simulation".  My question is whether or not you measure
orthogonality (sp?) of the design on the basis of the corresponding
attribute arrays or on the differences between the two sets of arrays.

For example, if you have attribute X with levels 100, 200, 300 -- in one
alternative you have X = 100 and in the other X=300.  Do you measure the
orthogonal foundation of your design on the basis of the two attribute X
measures or on the basis of the difference in the X measures.

The second, and related, question then becomes the necessary number of
pair scenarios in order to estimate the main effects model,  If you have
5 attributes in each alternative, do you need to determine the number of
choice scenarios on the basis of the 10 attributes (5 attributes X 2
alternatives) or on the basis of the 5 difference measures.  Some of the
materials I read say 5, others say 10.  Presuming the choice model will
be estimated using MNL regression.

Thanks Again.

Rick Perdue

Rick Perdue

Professor and Department Head
Department of Hospitality and Tourism Management
Virginia Tech
Blacksburg, VA 24061
540-231-5515
FAX [phone]

Editor:  Journal of Travel Research
President:  International Academy for the Study of Tourism

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Anthony Babinec
Sent: Wednesday, October 03, 2007 9:24 PM
To: [hidden email]
Subject: Re: Paired Choice Fractional Experiment Design Question

ORTHOPLAN was designed with rank- or ratings-based conjoint
analysis in mind, and not choice modeling. However, you should
be able to make ORTHOPLAN work for you by doubling your number
of columns - that is, generate an orthogonal design that has columns

att1_alt1 att2_alt1 ... att5_alt1 att1_alt2 att2_alt2 ...att5_alt2

For more on the use of ORTHOPLAN in choice models, see "Applied
Choice Analysis" by David Hensher, John Rose, and William Greene,
published by Cambridge University Press.

Anthony Babinec
[hidden email]

"Be the change you want to see in the world." Gandhi

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Perdue, Richard
Sent: Wednesday, October 03, 2007 12:02 PM
To: [hidden email]
Subject: Paired Choice Fractional Experiment Design Question

Hi Folks



My apologies if you've talked about this in the past.  I'm a newcomer to
this list.



I am doing a project examining consumer preferences and would like to
use a fractional factorial design to do a paired choice preference
study.

I have five attributes with (4)(2)(4)(2)(4) levels.  I am able to use
the SPSS design function to create an orthogonal, main effects only
model for one presentation of these attributes, but am uncertain how to
do the paired presentation.



Your advice would be greatly appreciated.



Thanks



Rick Perdue

Virginia Tech

[Anthony Babinec]

Richard,
My reading of Hensher et al. suggests that you are interested in the overall
array, and not in differences between the two sets.

Regarding degrees of freedom, Hensher et al.'s Table 5.10 presents some
formulas for the minimum number of treatment combinations required for main
effects fractional factorial designs. The question for you is whether you
are interested in linear effects only or in general effects. A complicating
factor in your design is that your attributes do not have the same number of
levels.

With linear effects only, the minimum number of treatment combinations would
be 2*5+1 or 11.

With general effects and all attributes at 4 levels, the minimum number of
treatment combinations would be (4-1)*2*5+1 equals 31.

It appears that Orthoplan in fact generates 64 "cards" when all attributes
are 4 levels. If I change two of the attributes to 2 levels, then Orthoplan
generates a design with 32 cards. The minimum number of cards needed is less
than this, but I don't think Orthoplan will do anything different for this
design.

For any design that Orthoplan generates, you should inspect the pairs being
compared. I don't think that there is any way to constrain the design, for
example, declare infeasible points. You can of course regenerate a design.
To invoke restrictions or generate a design with a fewer number of trials,
you need to look elsewhere than Orthoplan, such as Sawtooth Software's CBC
or SAS's experimental design generator.

Whatever number of trials you use, you might want to mock up some data to
verify that you can estimate the choice model.



Anthony Babinec
[hidden email]

"Be the change you want to see in the world." Gandhi
-----Original Message-----
From: Perdue, Richard [mailto:[hidden email]]
Sent: Wednesday, October 03, 2007 9:10 PM
To: Anthony Babinec; [hidden email]
Subject: RE: Paired Choice Fractional Experiment Design Question

Thanks Anthony

I have the Hensher, Rose and Greene book as well as the one by Louviere,
Hensher and Swait "Stated Choice Methods" and Train's "Discrete Choice
Methods with Simulation".  My question is whether or not you measure
orthogonality (sp?) of the design on the basis of the corresponding
attribute arrays or on the differences between the two sets of arrays.

For example, if you have attribute X with levels 100, 200, 300 -- in one
alternative you have X = 100 and in the other X=300.  Do you measure the
orthogonal foundation of your design on the basis of the two attribute X
measures or on the basis of the difference in the X measures.

The second, and related, question then becomes the necessary number of
pair scenarios in order to estimate the main effects model,  If you have
5 attributes in each alternative, do you need to determine the number of
choice scenarios on the basis of the 10 attributes (5 attributes X 2
alternatives) or on the basis of the 5 difference measures.  Some of the
materials I read say 5, others say 10.  Presuming the choice model will
be estimated using MNL regression.

Thanks Again.

Rick Perdue

[Perdue, Richard]

Thanks Anthony

Your conclusions match with mine.  Always a good feeling!

I've been able to create an efficient, orthogonal plan with 32 choice
scenarios.  Both sets of scenarios are orthogonal as is the overall set.
When I calculate the differences between the pairs, the correlations
between the attribute difference scores is less than 0.1 in all cases.
I will set it up as two blocks of 12 and one of 11 with a throw-away on
the front end of each.

I will be collecting pilot data next week to make sure that the model
can be estimated.

Thanks again.

Rick Perdue

Rick Perdue

Professor and Department Head
Department of Hospitality and Tourism Management
Virginia Tech
Blacksburg, VA 24061
540-231-5515
FAX [phone]

Editor:  Journal of Travel Research
President:  International Academy for the Study of Tourism

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Anthony Babinec
Sent: Thursday, October 04, 2007 11:35 AM
To: [hidden email]
Subject: Re: Paired Choice Fractional Experiment Design Question

Richard,
My reading of Hensher et al. suggests that you are interested in the
overall
array, and not in differences between the two sets.

Regarding degrees of freedom, Hensher et al.'s Table 5.10 presents some
formulas for the minimum number of treatment combinations required for
main
effects fractional factorial designs. The question for you is whether
you
are interested in linear effects only or in general effects. A
complicating
factor in your design is that your attributes do not have the same
number of
levels.

With linear effects only, the minimum number of treatment combinations
would
be 2*5+1 or 11.

With general effects and all attributes at 4 levels, the minimum number
of
treatment combinations would be (4-1)*2*5+1 equals 31.

It appears that Orthoplan in fact generates 64 "cards" when all
attributes
are 4 levels. If I change two of the attributes to 2 levels, then
Orthoplan
generates a design with 32 cards. The minimum number of cards needed is
less
than this, but I don't think Orthoplan will do anything different for
this
design.

For any design that Orthoplan generates, you should inspect the pairs
being
compared. I don't think that there is any way to constrain the design,
for
example, declare infeasible points. You can of course regenerate a
design.
To invoke restrictions or generate a design with a fewer number of
trials,
you need to look elsewhere than Orthoplan, such as Sawtooth Software's
CBC
or SAS's experimental design generator.

Whatever number of trials you use, you might want to mock up some data
to
verify that you can estimate the choice model.



Anthony Babinec
[hidden email]

"Be the change you want to see in the world." Gandhi
-----Original Message-----
From: Perdue, Richard [mailto:[hidden email]]
Sent: Wednesday, October 03, 2007 9:10 PM
To: Anthony Babinec; [hidden email]
Subject: RE: Paired Choice Fractional Experiment Design Question

Thanks Anthony

I have the Hensher, Rose and Greene book as well as the one by Louviere,
Hensher and Swait "Stated Choice Methods" and Train's "Discrete Choice
Methods with Simulation".  My question is whether or not you measure
orthogonality (sp?) of the design on the basis of the corresponding
attribute arrays or on the differences between the two sets of arrays.

For example, if you have attribute X with levels 100, 200, 300 -- in one
alternative you have X = 100 and in the other X=300.  Do you measure the
orthogonal foundation of your design on the basis of the two attribute X
measures or on the basis of the difference in the X measures.

The second, and related, question then becomes the necessary number of
pair scenarios in order to estimate the main effects model,  If you have
5 attributes in each alternative, do you need to determine the number of
choice scenarios on the basis of the 10 attributes (5 attributes X 2
alternatives) or on the basis of the 5 difference measures.  Some of the
materials I read say 5, others say 10.  Presuming the choice model will
be estimated using MNL regression.

Thanks Again.

Rick Perdue

[Anthony Babinec]

Richard, I had to read Hensher et al over and over
to feel that I grasped it. Sounds like you are on the
right track.

Anthony Babinec
[hidden email]

"Be the change you want to see in the world." Gandhi

-----Original Message-----
From: Perdue, Richard [mailto:[hidden email]]
Sent: Thursday, October 04, 2007 11:21 AM
To: Anthony Babinec; [hidden email]
Subject: RE: Paired Choice Fractional Experiment Design Question

Thanks Anthony

Your conclusions match with mine.  Always a good feeling!

I've been able to create an efficient, orthogonal plan with 32 choice
scenarios.  Both sets of scenarios are orthogonal as is the overall set.
When I calculate the differences between the pairs, the correlations
between the attribute difference scores is less than 0.1 in all cases.
I will set it up as two blocks of 12 and one of 11 with a throw-away on
the front end of each.

I will be collecting pilot data next week to make sure that the model
can be estimated.

Thanks again.

Rick Perdue

Rick Perdue

Professor and Department Head
Department of Hospitality and Tourism Management
Virginia Tech
Blacksburg, VA 24061
540-231-5515
FAX [phone]

Editor:  Journal of Travel Research
President:  International Academy for the Study of Tourism