Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Experimental Designs for CBC
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Outline
2
Background – designs for linear models
CBC design strategies
Based on orthogonal arrays
Using efficiency calculations
Sawtooth Software’s multi-objective optimization
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
BACKGROUND
3
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
In the absence of an experimental design bad and terrible things can happen:
Terrible: Some of your model coefficients (utilities) may not be possible to estimate, your research becomes a crash-and-burn disaster, dollars and careers can go up in flames
Bad: your utility estimates lose precision
If you don’t get precision through your experimental design, you have to buy it with sample size
Savings in $, €, ¥, £, Rs are good reasons to like precision
Why care?
4
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Attributes are the characteristics of products and levels are the discrete values those characteristics may take on
Attribute (levels):
Brand (Pepsi, Coke, Dr. Pepper)
Capacity (16GB, 32GB, 64GB)
Price ($90, $125, $159)
Flame-retardant fabric (yes, no)
Attributes and Levels
5
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Orthogonal designs
“Full factorial” designs
Show every possible combination of predictors
An experiment with three 4 level factors and two 5 level factors (43 x 52) would need observations for 4 x 4 x 4 x 5 x 5 = 1,600 combinations
Fractional factorial plans require fewer observations
They assume no or very few interactions
When they measure main effects only, we call them orthogonal main effects plans (OMEPs)
An OMEP for the (43 x 52) design above uses just 25 combinations or “runs”
Design for Linear Models
6
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
34 OMEP
Example Design
7
Run X1 X2 X3 X4
1 3 3 3 3
2 3 1 1 2
3 3 2 2 1
4 1 3 1 1
5 1 1 2 3
6 1 2 3 2
7 2 3 2 2
8 2 1 3 1
9 2 2 1 3
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Orthogonality - each level occurs with levels from each other attribute a proportional number of times
Correlation matrix:
Two-way frequencies:
Properties of Example
8
X1 X2 X3 X4
X1 1
X2 0 1
X3 0 0 1
X4 0 0 0 1
X1
0 1 2
X2
0 1 1 1
1 1 1 1
2 1 1 1
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Addelman (1962) OMEP catalog
Designs account for main effects (of individual attributes) not interactions
Until 1994, the marketing researcher’s bible for creating experimental designs
Hahn and Shapiro (1966) designs allowed some interactions
Kuhfeld’s SAS Technical Report “Orthogonal Arrays” offers an extensive collection plus links to other design databases
Sources of Orthogonal Fractional Designs
9
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Traditional conjoint analysis used fractional factorial experimental designs to create a set of stimuli (concepts, profiles, cards)
Respondents evaluated these stimuli via ratings or rankings
Regression analysis decomposed the evaluations into utilities for each level of each attribute in the study
We can also use fractional factorials in
Menu-based choice experiments
Situational choice experiments
Choice-based conjoint (CBC) experiments
Using Orthogonal Arrays
10
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
CBC DESIGNS FROM ORTHOGONAL
PLANS
11
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Adapting fractional factorial designs for CBC experiments
Louviere and Woodworth 1983 – 2J designs
4 more early strategies (Louviere, Hensher and Swait, 2000)
Mix and match 1
Mix and match 2
Labeled mix and match
LMN
Simple shifting (Bunch, Louviere and Anderson 1996)
Flexible shifting (Street and Burgess 2007, Street, Burgess and Louviere 2005)
Designs Based on Orthogonal Plans
12
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
1. Use a fractional factorial design to create profiles
2. Use a second fractional factorial design to decide which profiles to assign to particular choice sets (levels are “absent” and “present”
3. Hadamard matrices or balanced incomplete block designs can be used for step 2 as well
2J Designs
13
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Pros
Conceptually and mechanically simple
Cons
Designs can get very large
Different choice sets can have different numbers of profiles
Pros and Cons
14
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Mix and Match 1
15
Algorithm
1. Generate a fractional factorial plan
2. Make some number of k exact copies of this fraction
3. From each of the k+1 equivalent fractions
Randomly select one profile without replacement
This set of K+1 profiles is choice set 1
Prevent duplicates within a set
4. Repeat step 3 until all profiles are assigned to choice sets
This strategy produces an experiment where all choice sets have k+1 profiles
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
The same 9 profiles appear in both the first and second alternative
Example of Mix and Match 1
16
Alternative A Alternative B
Set X1 X2 X3 X4 X1 X2 X3 X4
1 3 3 3 3 1 3 1 1
2 3 1 1 2 2 2 1 3
3 3 2 2 1 2 1 3 1
4 1 3 1 1 1 2 3 2
5 1 1 2 3 3 2 2 1
6 1 2 3 2 3 3 3 3
7 2 3 2 2 1 1 2 3
8 2 1 3 1 2 3 2 2
9 2 2 1 3 3 1 1 2
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Mix and Match 2
17
Algorithm
1. Generate a fractional factorial plan
2. Make some number of k equivalent (not exact) copies of this fraction*
3. From each of the k+1 equivalent fractions
Randomly select one profile without replacement
This set of K+1 profiles is choice set 1
4. Repeat step 3 until all profiles are assigned to choice sets
This strategy covers more of the design space than Mix and Match 1
*You can do this simply by switching columns, relabeling levels (1 2, 23 and 31) or both.
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
In this experiment unique profiles appear in the two alternatives and we cover twice as many possible profiles
Example of Mix and Match 2
18
Alternative A Alternative B
Set X1 X2 X3 X4 X1 X2 X3 X4
1 3 3 3 3 2 3 2 1
2 3 1 1 2 1 2 3 1
3 3 2 2 1 1 1 2 2
4 1 3 1 1 2 1 3 3
5 1 1 2 3 3 1 1 1
6 1 2 3 2 1 3 1 3
7 2 3 2 2 2 2 1 2
8 2 1 3 1 3 3 3 2
9 2 2 1 3 3 2 2 3
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
We could create multiple versions of a mix and match design
One could be the design on the previous slide
One could be the previous design where columns 1&2 and 3&4 are swapped in the left hand profile
One could swap codes 3 and 2 in the even numbered columns of the right hand profile and 1 and 2 in the odd columns of the right hand profile
We could make several other versions in a similar way, swapping columns and/or recoding levels
Then we could randomly assign different respondents to different versions
When we do this we explore more of the “design space”
We also allow estimation of some interactions
Versions
19
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
A trick called “labeling” allows us to make alternative-specific designs In each choice set, label the profile chosen from the first of the K+1 designs
with the first brand name
Label the profile chosen from the second design with the second brand name
Etc.
Benefits Brand is now an attribute in your experiment that you didn’t have to build
into the fractional factorial experiment
In analysis, you can have brand-specific effects of price and all other attributes
In other words you can estimate interactions of brand with all other attributes
Alternative-Specific Designs
20
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Just label the sets of alternatives with brand names
Example of Alternative-Specific Design
21
Bose JBL
Set X1 X2 X3 X4 X1 X2 X3 X4
1 3 3 3 3 2 3 2 1
2 3 1 1 2 1 2 3 1
3 3 2 2 1 1 1 2 2
4 1 3 1 1 2 1 3 3
5 1 1 2 3 3 1 1 1
6 1 2 3 2 1 3 1 3
7 2 3 2 2 2 2 1 2
8 2 1 3 1 3 3 3 2
9 2 2 1 3 3 2 2 3
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
LMN Designs
22
If we have L levels per attribute, M alternatives in our choice sets and N attributes, build the design from a fractional factorial with M*N columns
For example, if we have 4 alternatives with 3 attributes of 3 levels each we need a design with 12 columns worth of 3 level attributes – a 27 row design from the Addelman catalog would work
In this design variables are independent both within and across profiles
This feature allows measurement of cross-effects
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
3 alternatives, 2 attributes of 3 levels each - this design uses an 18 run fractional factorial design with 7 columns of 3-level variables – we use two per alternative, with one unused column
Note we have an unused column. We need not use every column of a design. We may also use it for other purposes, like creating “blocks” of questions so that no respondent has to answer all 18.
A 33*2 Design
23
Set A1 A2 B1 B2 C1 C2 Unused
1 3 3 3 1 2 3 3
2 3 1 1 2 1 1 2
3 3 2 2 3 3 2 1
4 1 3 1 3 1 2 3
5 1 1 2 1 3 3 2
6 1 2 3 2 2 1 1
7 2 3 2 2 1 3 1
8 2 1 3 3 3 1 3
9 2 2 1 1 2 2 2
13 3 3 2 3 2 1 2
11 3 1 3 1 1 2 1
12 3 2 1 2 3 3 3
13 1 3 3 2 3 2 2
14 1 1 1 3 2 3 1
15 1 2 2 1 1 1 3
16 2 3 1 1 3 1 1
17 2 1 2 2 2 2 3
18 2 2 3 3 1 3 2
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Shifting Designs
24
1. Generate a fractional factorial plan
2. Results of this plan are the first profile of each choice set
3. Second profile is just the first with all attributes “shifted”
1 becomes 2, 2 becomes 3 so on, with wraparound (highest level becomes 1 again)
Example:
Six 3-level attributes
If first profile is 331232
Second profile is 112313 and third is 223121
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Starting with a 34 fractional factorial we create additional alternatives through shifting
Note there is no level overlap
Note that a 4th profile would duplicate the first
Shifting Design Example
25
Alternative A Alternative B Alternative CSet X1 X2 X3 X4 X1 X2 X3 X4 X1 X2 X3 X4
1 3 3 3 3 1 1 1 1 2 2 2 22 3 1 1 2 1 2 2 3 2 3 3 13 3 2 2 1 1 3 3 2 2 1 1 34 1 3 1 1 2 1 2 2 3 2 3 35 1 1 2 3 2 2 3 1 3 3 1 26 1 2 3 2 2 3 1 3 3 1 2 17 2 3 2 2 3 1 3 3 1 2 1 18 2 1 3 1 3 2 1 2 1 3 2 39 2 2 1 3 3 3 2 1 1 1 3 2
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
We could choose to shift different columns of the initial matrix differently
For example, we could shift the 1st and 3rd columns by one place and the 2nd and 4th by 2 places
Street and Burgess (2007) call this pattern of shifts, 1212, a “generator”
They show how to use generators to create a variety of designs for main effects and interactions
Like simpler shifting designs, these will not work for alternative-specific effects
Flexible Shifting Designs
26
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
This design uses a 1212 generator
Flexible Shifting Example
27
Alternative A Alternative B Alternative C
Set X1 X2 X3 X4 X1 X2 X3 X4 X1 X2 X3 X4
1 3 3 3 3 1 2 1 2 2 1 2 1
2 3 1 1 2 1 3 2 1 2 2 3 3
3 3 2 2 1 1 1 3 3 2 3 1 2
4 1 3 1 1 2 2 2 3 3 1 3 2
5 1 1 2 3 2 3 3 2 3 2 1 1
6 1 2 3 2 2 1 1 1 3 3 2 3
7 2 3 2 2 3 2 3 1 1 1 1 3
8 2 1 3 1 3 3 1 3 1 2 2 2
9 2 2 1 3 3 1 2 2 1 3 3 1
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
For generic main effects models, shifting and modified shifting give more efficient designs than do mix and match designs
Mix and match designs with enough versions allow estimation of interactions
For models with alternative-specific effects, use Mix and Match 2 or and LMN designs
For models with cross-effects, use LMN designs
Avoid shifting if you want to estimate interactions or alternative-specific effects
Shifting designs cause minimal overlap
Comparison of Orthogonal Designs
28
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Recipe-based design algorithms easy enough that you can create designs by hand or using simple Excel workbooks
While the columns of a fractional factorial are orthogonal with respect to one another, they may be perfectly correlated with un-modeled interactions
Focus on orthogonality can detract from efficiency
Prohibitions must be added manually and there is no good way to do this
Pros and Cons of Using Orthogonal Plans
29
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
EFFICIENT DESIGNS
30
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
We can calculate D-efficiency of a design mathematically
Rather than retrofitting orthogonal designs to create our CBC designs, we could search for groups of choice sets that constitute efficient designs
Computer search algorithms can assist this process
The covariance matrix itself depends on the model we plan to run, so different models require different efficiency calculations
Designs Based on Efficiency
31
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Kuhfeld, Tobias and Garratt (1994) show how to use search algorithms to make efficient designs for MNL models
Generate many sets of profiles with the right number of attributes and levels and let a search algorithm identify the most efficient way to combine them
SAS macros MKTEX and CHOICEFF work together to create a wide range of CBC experimental designs (Kuhfeld and Wurst 2012)
D-Efficiency for MNL Models
32
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
You can further improve designs with iterative refinements (Huber and Zwerina 1996) called relabeling and swapping
If you have reasonably good estimates of respondents’ utilities, you can do better still (Sandor and Wedel 2001)
These are sometimes called Dp-efficient designs because we have a guess about their parameters
They are distinguished from D0-efficient designs that assume null parameters
Further Improvements for MNL Models
33
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Efficient designs for mixed logit models (Sandor and Wedel 2002)
Efficient designs for hierarchical Bayesian models (Sandor and Wedel 2005)
Efficient designs for two-stage models with a non-compensatory first stage (Liu and Arora 2011)
Efficient designs for repeated measures logit models (Bliemerand Rose 2010)
Other Model Specifications
34
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Relatively easy to implement the basic D0-efficient MNL model (using off-the-shelf software)
Complexity grows rapidly for different models and requires
Extreme programming skills and/or
Specialized software and/or
Prior estimates of utilities
Pros and Cons of Efficiency-Based Designs
35
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Efficiency based designs are more flexible than orthogonal designs (e.g. prohibitions are easier to accommodate)
Efficiency-based designs by their nature will be a little more efficient than orthogonal designs - a lot more efficient according to Bliemer and Rose (2010)
The differences may not matter much in practice
A lot of sniping and grousing happens anyway
Orthogonal versus Efficiency-Based Designs
36
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
LIGHTHOUSE STUDIO DESIGNS
37
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
These properties promote efficiency
(One-way) Level Balance: within each attribute, each level appears an equal number of times
Orthogonality, or two-way level balance: each level appears proportionally often with every level of other attributes
Minimal Overlap: minimizing the extent to which an attribute level repeats within a choice set
Sources of CBC Efficiency
38
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
We can balance the sources of efficiency to produce four design strategies
Good designs at the respondent level that get better as more respondents answer more versions of the design
As the number of versions increases, the difference in D-efficiency between complete enumeration and efficient designs becomes minuscule (Kuhfeld and Wurst 2012)
Using many versions means
We can cover a larger portion of the design space
We can estimate interactions without pre-planning for them
Version effects can cancel out
Lighthouse Studio’s Design Strategies
39
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Generation process:
Maximize one- and two-way level balance
Profiles are nearly orthogonal for a given respondent
Features
One-way level balance (each level of an attribute occurs as equally often as possible)
Two-way level balance (each pair of levels occurs as equally often as possible)
Minimal overlap of levels within choice sets (promotes precision of main effects)
Complete Enumeration
40
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Generation process: Similar to Complete Enumeration, except:
Profiles are built using previously least-used attribute levels (from previous tasks and versions)
Two-way frequency balance is not explicitly controlled
Features
Minimal overlap of levels between profiles
One-way level balance (each level of an attribute occurs as equally often as possible)
Loose (not strict) orthogonality
Shortcut
41
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Generation process
Profiles sampled randomly, with replacement, from full factorial universe of profiles
No duplicate profiles allowed within a choice set
Features
Only loose one-way level balance (each level occurs approximately an equal number of times)
Only loose two-way level balance (each pair of levels occurs approximately an equal number of times)
Significant level overlap (an advantage for precision of interactions)
Random
42
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Similar to Complete Enumeration but with more level overlap
Features
About half as much level overlap as the purely Random strategy
One-way level balance (each level occurs an equal number of times)
Two-way level balance (each pair of levels occurs an equal number of times)
Much better for main effects than Random and much better for interaction effects than Complete Enumeration
Balanced Overlap is the default method in Sawtooth Software’s CBC software
Balanced Overlap
43
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
CompleteEnumeration
Shortcut Random Balanced Overlap
Fast + +
Main Effect Efficiency + + -
Interaction Efficiency - - + +
Deep Processing + +
Extreme Prohibitions + +
Pros and Cons of Lighthouse Strategies
44
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
QUESTIONS?
45
Megan PeitzIngenuity Ambassador
Keith ChrzanSVP, Sawtooth Analytics
www.sawtoothsoftware.com+1 801 477 4700@sawtoothsoft
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Addelman, S. (1962) “Orthogonal Main-effects Plans for Asymmetrical Factorial Experiments, Technometrics, 4: 21-46.
Bliemer, M.C.J. and J.M. Rose (2010) “Construction of Experimental Designs for Mixed Logit Models Allowing for Corelation Across Choice Observations,” Transportation Research Part B, 44, 720-734.
Bliemer, M,C.J., J.M. Rose and S. Hess (2008) “Approximation of Bayesian Efficiency in Experimental Choice Designs,” Journal of Choice Modeling, 1:98-127.
Bunch, D.S., J.J. Louviere and D. Anderson (1996) “A Comparison of Experimental Design Strategies for Multinomial Logit Models: The Case of Generic Attributes,” accessed on January 26, 2013 at http://faculty.gsm.ucdavis.edu/~bunch/bla_wp_1996.pdf.
Chrzan, K. and B. Orme (2000) “Overview and Comparison of Design Strategies for Choice-Based Conjoint Analysis,” Sawtooth Software Conference Proceedings, 161-177)
Hahn GJ. Shapiro SS. (1966) A catalogue and computer program for the design and analysis of orthogonal symmetric and asymmetric fractional factorial experiments. Report 66-C-165, General Electric Research and Development Centre, Schenectady, New York.
References
46
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Huber, J. and K. Zwerina (1996) “The Importance of Utility Balance in Efficient Choice Designs,” Journal of Marketing Research, 33: 307-317.
Kanninen, B. (2002) “Optimal Designs for Multinomial Choice Experiments,” Journal of Marketing Research, 39: 214-227.
Kuhfeld, W.F. Orthogonal Arrays. Technical Report, SAS Institute. Accessed on January 26, 2013 at http://support.sas.com/techsup/technote/ts723.html.
Kuhfeld, W.F, R. Tobias and M. Garratt (1994) “Efficient Experimental Design with Marketing Research Applications,” Journal of Marketing Research, 31, 545-557.
Kuhfeld, W.F. and J.C. Wurst (2012) “An overview of the Design of Stated Choice Experiments,” Sawtooth Software Conference Proceedings. Orem: Sawtooth Software
Liu, Q. and N. Arora (2011) “Efficient Choice Designs for a Consider-Then-Choose Model,” Marketing Science, 30: 321-338.
Louviere, J.J., D.A. Hensher and J.D. Swait (2000) Stated Choice Methods: Analysis and Application. Cambridge: Cambridge University
References
47
© 2016 Sawtooth Software, Inc. | www.sawtoothsoftware.com
Webinar
Louviere, J.J. and G.W. Woodworth (1983) “Design and Analysis of Simulated Consumer Choice or Allocation Experiments: An Approach Based on Aggregate Data,” Journal of Marketing Research, 20: 350-367.
Sandor, Z. and M. Wedel (2001) “Designing Choice Experiments Using Managers’ Prior Beliefs,” Journal of Marketing Research, 38: 430-444.
Sandor, Z. and M. Wedel (2002) “Profile Construction in Experimental Choice Designs for Mixed Logit Models,” Marketing Science, 21: 455-475.
Sandor, Z. and M. Wedel (2005) “Heterogeneous Choice Designs,” Journal of Marketing Research, 42: 210-218.
Street, D.J. and L. Burgess (2007) The Construction of Optimal Stated Choice Experiments. Hoboken: Wiley.
Street , D.J., L. Burgess and J.J. Louviere (2005) “Quick and Easy Choice Sets: Constructing Optimal and Nearly Optimal Stated Choice Experiments,” International Journal of Research in Marketing, 22: 459-470.
References
48