webinar - a beginners guide to choice-based conjoint analysis
TRANSCRIPT
A Beginner’s Guide toChoice-based Conjoint
AnalysisPaul Richard McCullough & Ray Poynter
Agenda
1. Introduction2. What is Choice-based Conjoint
Analysis?3. Key steps in Conducting Choice-based
studies4. Presenting the results5. Critical elements for success6. Q & A 2
Ray PoynterMD The Future
Place
Paul Richard McCullough President, MACRO Consulting
3
Simple Example – Flying
Return flight journey – 2 to 3 hours
3 Simple AttributesAirline: Shabby, OK, ClassyCost: $75, $150, $225Route: Direct, 1 change, 2
changes 4
Task 1
Classypt OK Shabby$225 $75 $150Two
changes Direct One change☐ ☐ ☐
5
Task 1
Classy OK Shabby$225 $75 $150Two
changes Direct One change☐ ☐
PickedShabby 0OK 1Classy 0$75 1$150 0$225 0Direct 1One change 0Two changes 0
Task 2
Classy OK Shabby$75 $150 $225
Direct One changeTwo
changes ☐ ☐
PickedShabby 0OK 1Classy 1$75 2$150 0$225 0Direct 2One change 0Two changes 0
Task 3
Classy OK Shabby$150 $225 $75
One Change DirectTwo
Changes ☐ ☐
PickedShabby 0OK 1Classy 2$75 2$150 1$225 0Direct 2One change 1Two changes 0
Task 3
Classy OK Shabby$150 $225 $75
One Change DirectTwo
Changes ☐ ☐
Picked % of 3Shabby 0 0%OK 1 33%Classy 2 67%$75 2 67%$150 1 33%$225 0 0%Direct 2 67%One change 1 33%Two changes 0 0%
Utilities After 11 Tasks
Picked OfferedTransfor
mImportan
ceShabby 2 18%OK 4 36%Classy 5 45%$75 6 55%$150 4 36%$225 1 9%Direct 7 64%One change 3 27%Two changes 1 9%
Utilities After 11 Tasks
Picked OfferedTransfor
mImportan
ceShabby 2 18% 0OK 4 36% 33Classy 5 45% 50$75 6 55% 83$150 4 36% 50$225 1 9% 0Direct 7 64% 100One change 3 27% 33Two changes 1 9% 0
Utilities After 11 Tasks
Picked OfferedTransfor
m
RelativeImportan
ceShabby 2 18% 17
21%OK 4 36% 50Classy 5 45% 67$75 6 55% 83
36%$150 4 36% 50$225 1 9% 0Direct 7 64% 100
43%One change 3 27% 33Two changes 1 9% 0
Basic Concepts And Vocabulary
Attributes are variables that define the productLevels are variable values Attribute Levels
Brand CokePepsi
7upPrice 50¢
$1.00 $1.50
Twist Off Cap IncludedNot Included
13
Alternatives and Tasks
Combinations of attribute levels define alternatives:
Alternative #1: Pepsi @ $1.00 with a twist off cap
A choice task is a set of alternatives14
Choice-based Conjoint Task
15
Utility Weights
Utility weights are a measure of how important an attribute level is to the choice decision
Utility weights are analogous to beta weights, aka, regression coefficients, in a regression model
16
Partworths and VectorsPartworth attributes are categorical, e.g., brand,
color, package designEach level within a partworth attribute has a
utility weightVector attributes are interval or ratio, e.g., price,
CPU speed, horsepowerVector attributes receive one utility weight for
the entire attribute (all levels)17
Partworths and Vectors
18
Attribute Relative Importance
Attribute Relative Importance is the difference between high and low attribute level utility weights, ie, gap or delta, percentaged against all other attribute deltas
It’s used almost universallyAnd it’s often misleading
19
Attribute Relative Importance
Attribute Levels Util Delta ImportanceBrand Coke 6 5 50%
Pepsi 27up 1
Price 50¢ 7 4 40%$1 5
$1.5 3Twist Off Cap Included 0.5 1 10%
Not Included -0.5
20
What Happens When I Change Price?
Attribute Levels Util Delta ImportanceBrand Coke 6 5 63%
Pepsi 27up 1
Price 75¢ 6 2 25%$1 5
$1.25 4Twist Off Cap Included 0.5 1 13%
Not Included -0.5
21
When To Use Conjoint?❏The value can be approximated from the sum of the parts White goods, computers, financial products, pack & size combinations for CPG, X Soft drinks, clothing, TV shows, political programs, life partners
❏Attributes do not interact Credit card features, mobile phone packages, car rental options, ski holiday
elementsX Fragrances, speeches, meal ingredients, newspaper front page designs, advertising
messages
❏Attributes can be described in discrete and unambiguous terms
Apple iOS or Android, 2 or 4 door, $100 or $200 or $300, glass or plastic bottleX Friendly, Clean, Quick, Sexy, Sporty, Ambitious, Relaxed
22
Choice-based Conjoint – Key Steps
❏Context- What is the business issue we want to answer?❏Attributes and levels- How do we define the alternatives in
terms of attributes and levels?
❏Model specification- Which form of CBC is best?
❏Analysis- Which analytic approach is best?
23
Context
CBC is about choiceBut which choice?Purchase is most common choice measuredThere are others:• Subscribe to a magazine• Ask someone for a date• Rent a movie for yourself and your significant other• Order dinner• Consume electricity in your yacht• Plan a vacation for the familyBe clear to set the context for the choice at the beginning of the exercise
24
Case Study
Manufacturer of Safety GlassesHighly regulated environmentPurchasers are experienced and knowledgeable about
categoryChoice in this context is purchase of safety glasses for
company
25
Attributes and Levels
Brand (5 brands)Style (8 styles graphically displayed)Price (5 price levels)Origin:
Made in the USA Made in Asia
Included/Not included:1. Adjustable nose bridge2. Fully adjustable3. Ratcheting temples4. Multiple Sizes5. Soft temple Materials6. Low-Weight7. Flexible Fit (self adjust)8. Coating - Anti-scratch9. Coating - Anti-Fog10. Coating - Extreme Anti-Fog11. Coating - Anti Reflective12. Coating - Easy to Clean13. Coating - No coating
26
Choice-based Conjoint
a.k.a. Discrete Choice Models (DCM) Respondent chooses one or none from a set of
choices He/she sees several choice sets Based on a logit model, not linear regression
model Historically done at aggregate level
27
Discrete Choice ModelsPro’s
More realistic data collection (but not always) More accurate price and market share estimates
Con’s Limited number of attributes (unless more advanced form is
used) Less information per respondent Requires more knowledge to execute 28
Safety Glasses CBC Specifications17 attributes, 46 levels12 choice tasks4 alternatives per taskTraditional none option500 choice task set versionsSample size = 500
29
30
Choice Task
Other Design ConsiderationsLarge Number of Attributes: Full profile trade-off uses all attributes in each tested product
configuration Partial profile trade-off uses a subset of all attributes in each tested
product configuration Adaptive Methods generally attempt to eliminate or diminish
undesirable attributes prior to the conjoint tasks Hybrid Methods combine multiple exercises
31
Heterogeneity
What the heck is that?Half your sample loves CokeHalf your sample loves PepsiAn aggregate model will have a zero coefficient for brandDisaggregate models capture individual level differences,
ie, heterogeneity32
Estimating Disaggregate Choice Utilities
Ratings-based conjoint provides enough information to estimate individual level models for each respondent
Choice-based conjoint models typically do not
Two ways to estimate disaggregate CBC models: Latent Class Logistic Regression- Searches for homogeneous
subgroups Hierarchical Bayes Estimation- Estimates individual level
coefficients and draws from the total sample when needed33
Safety Glasses CBC Specifications
17 attributes, 46 levels12 choice tasks4 alternatives per taskTraditional none option500 choice task set versionsSample size = 500Full ProfileNull prohibitions (5-7)Hierarchical Bayes utility estimationLatent class choice utility estimation (for segmentation)
34
The Results
❏ Utilities❏ What-if
Models❏ Forecasting
35
Utilities
Picked OfferedTransfor
m
RelativeImportan
ceShabby 2 18% 0
21%OK 4 36% 33Classy 5 45% 50$75 6 55% 83
36%$150 4 36% 50$225 1 9% 0Direct 7 64% 100
43%One change 3 27% 33Two changes 1 9% 0 36
Utilities
A B CShabby 2 18% -0.82462OK 4 36% 0.17414Classy 5 45% 0.67506$75 6 55% 1.17188$150 4 36% 0.17078$225 1 9% -1.33086Direct 7 64% 1.67235One change 3 27% -0.32794Two changes 1 9% -1.33064
37
Utilities
R1 R2 R3 R4 R5 AVG TRANSShabby -0.73 -1.17 -1.11 -0.86 -0.88 -0.95 0OK -0.33 -0.18 -0.25 -0.45 -0.50 -0.34 24Classy 0.18 0.63 0.45 0.22 0.49 0.39 53$75 1.74 0.98 1.31 1.52 1.09 1.33 100$150 0.39 0.56 0.09 0.60 0.75 0.48 67$225 -1.41 -1.28 -0.95 -1.41 -1.16 -1.24 0Direct 1.03 1.25 1.47 1.09 1.19 1.21 93One change 0.16 0.45 0.29 0.31 0.45 0.33 59Two changes -1.04 -1.23 -1.29 -1.02 -1.42 -1.20 0
38
What-if Models
R1 R2 R3 R4 R5Shabby -0.73 -1.17 -1.11 -0.86 -0.88OK -0.33 -0.18 -0.25 -0.45 -0.50Classy 0.18 0.63 0.45 0.22 0.49$75 1.74 0.98 1.31 1.52 1.09$150 0.39 0.56 0.09 0.60 0.75$225 -1.41 -1.28 -0.95 -1.41 -1.16Direct 1.03 1.25 1.47 1.09 1.19One change 0.16 0.45 0.29 0.31 0.45Two changes -1.04 -1.23 -1.29 -1.02 -1.42
A - Shabby/$75/1chng 1.17B - OK/$150/Drct 1.10Buy A
Simple First Choice Model
39
What-if Models
R1 R2 R3 R4 R5Shabby -0.73 -1.17 -1.11 -0.86 -0.88OK -0.33 -0.18 -0.25 -0.45 -0.50Classy 0.18 0.63 0.45 0.22 0.49$75 1.74 0.98 1.31 1.52 1.09$150 0.39 0.56 0.09 0.60 0.75$225 -1.41 -1.28 -0.95 -1.41 -1.16Direct 1.03 1.25 1.47 1.09 1.19One change 0.16 0.45 0.29 0.31 0.45Two changes -1.04 -1.23 -1.29 -1.02 -1.42
A - Shabby/$75/1chng 1.17 0.26B - OK/$150/Drct 1.10 1.63Buy A B
Simple First Choice Model
40
What-if Models
R1 R2 R3 R4 R5Shabby -0.73 -1.17 -1.11 -0.86 -0.88OK -0.33 -0.18 -0.25 -0.45 -0.50Classy 0.18 0.63 0.45 0.22 0.49$75 1.74 0.98 1.31 1.52 1.09$150 0.39 0.56 0.09 0.60 0.75$225 -1.41 -1.28 -0.95 -1.41 -1.16Direct 1.03 1.25 1.47 1.09 1.19One change 0.16 0.45 0.29 0.31 0.45Two changes -1.04 -1.23 -1.29 -1.02 -1.42
A - Shabby/$75/1chng 1.17 0.26 0.49 0.97 0.65B - OK/$150/Drct 1.10 1.63 1.31 1.25 1.44Buy A B B B B
Simple First Choice Model
41
Interactive Model ExampleDiscrete Choice Demonstration Model
42
What-if Models
❏Add, delete, modify products to test
❏Demographic sub-groups, e.g. male/female, users/non-users
❏Iterative test to identify optimum strategies
❏Add factors such as cost, revenue and profit
43
Forecasting
❏Converting models to the real world
❏Converting preference to predictions Use benchmarks or models
❏Adding in external effects Distribution, advertising, marketing,
inertia, etc44
Conjoint Challenges
❏Number of Levels Effect
❏IIA
❏Interactions
45
Number of Levels EffectStudies show that increase the number of levels & attribute becomes more importantFor example if we have 2 attributesColour: Red and BluePrice $1 & $2Might find that Colour accounted for 40% and Price for 60%
But if the options wereColour: Red and BluePrice $1, $1.50 & $2Might find Colour accounting for 30% and Price for 70%
46
IIA Problems
IIA - Independence from Irrelevant Alternatives Problems
Red Bus, Blue Bus problem
Respondent gives 50% to taxi & 50% to Blue BusIn this model the color has no transport value
Add Red Bus to the modelSame profile as Blue BusShares: Taxi 33%, Red Bus 33%, Blue Bus 33%
47
Interactions
In Conjoint theory attributes should be independent, i.e. they all have a single value, ignoring the other attributes they a combined with
However some attributes interact: You might like Iced Coke, and Hot Tea, and Iced Tea, but not Hot Coke – an
interaction
Two solutions: 1. Combine the attributes Hot tea, Iced Tea, Cold Coke2. Use more complex designs and analysis 48
Main Forms of Conjoint ModelsThere are four main classes of trade-off models:
Choice-based Conjoint-• Discrete Choice• Menu-based Conjoint- Respondents “build” their own productRatings-based Conjoint- Respondents rate one product profile at a timeSelf-Explicated Scaling-• Each attribute level is rated separately• Not really conjoint but is used in the same way, similar resultsHybrid models-Two different techniques are bridged together, usually CBC or ratings-based conjoint and self-explicated• Adaptive conjoint, eg, ACA or ACBC, are hybrid models• Adaptive models usually attempt to eliminate or diminish attributes prior to
the conjoint exercise
49
Choice-based Conjoint Task
50
Menu-based Conjoint
51
Ratings-based Conjoint
52
Self-Explicated Ratings
1
Adds no value
whatsoever 2 3 4 5 6 7 8 9
10 Adds a great
deal of value
200 mhz CPU speed nmlkj nmlkji nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj USB Port nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkji nmlkj nmlkj nmlkj Printer Port nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkji nmlkj Ethernet Port nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkji nmlkj nmlkj nmlkj 30 day satisfaction guarantee nmlkj nmlkj nmlkj nmlkj nmlkji nmlkj nmlkj nmlkj nmlkj nmlkj 300 mhz CPU speed nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkji 24/7 live phone tech support nmlkj nmlkj nmlkji nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj nmlkj Wireless keyboard nmlkj nmlkji nmlkj nmlkj nmlkj nmlkj nmlkj
nmlkj nmlkj nmlkj
2. Below are potential Network Computer features. Please rate each feature on how much value they add to a Network Computer assuming a rating of 10 means the feature adds a great deal of value and a rating of 1 means the feature adds no value whatsoever to the Network Computer system.
53
Presenting The Results1. The conjoint study is not the project, the
business problem being tackled is the project
2. The conjoint data / model is not the answer, a business action or solution will be the answerClients don’t want to hear how hard you worked, or how clever
you are.They want answers to their problems, not a lesson on your interests
54
Workshops Are Often Best
55
General Advice For Quant Presenting
❏ Minimize the use of numbers
❏ Minimize the number of digits, 1, 2, or occasionally 3 significant digits
❏ Avoid or minimize decimal places and negative numbers
❏ Make differences as large as possible – regroup if necessary
❏ Order and structure the information (e.g. rank by importance or value)
❏ Show comparisons
56
Bad and Not So Bad Examples
21.25%
40.67%
38.08%
Attribute Importance
Standard Price Changes
Price41%
Changes38%
Standard21%
Attribute Importance
57
Bad Example – Questionnaire Order
OK Shabby Classy $225 $150 $75 Direct One change
Two changes
-0.34
-0.95
0.39
-1.24
0.48
1.33 1.21
0.33
-1.20
Utilities
Better
$225 $150 $75 Two changes
One change
Direct Shabby OK Classy
0
67
100
0
60
94
0
24
52
Utilities
Critical Elements for Success
1. Understanding the business problem2. Including all of the relevant attributes and levels – e.g. the 0
level3. Create a survey that respondents can complete
Not too long, not confusing, suitable for the devices being used4. Creating a fixed context for each interview (e.g. breakfast menu)5. Attributes that do not interact & which are unambiguous6. Build a model early, 10 to 20 interviews, to check you are on
track7. Try to use a workshop as the debrief, rather than a report or
presentation8. Focus on business problem, not the conjoint methodology/model
60
Ray PoynterMD The Future Place
Paul Richard McCullough President, MACRO Consulting
@questionpro
www.questionpro.com
Questions?