Conjoint Analysis

1
Conjoint Analysis

• Session March 22-26 2010

2
1. Objectives/Purpose

• An extremely powerful and useful analysis tool
• Used to determine the relative importance of
  various attributes to respondents, based on their
  making trade-off judgments
• Useful in
• Helping to select features on a new
  product/service
• Predicting sales
• Understanding decision processes/consumer
  judgments

3
1. Objectives (ctd)

• E.g.
• UvT What drives students choice of (and
  willingness to pay for) a room?
• How can Albert Heijn compose its assortment of
  cereals to improve customer appeal?
• Nike What are the optimal features for a new
  type of sneakers?

4
2. Steps

• Design
• Assumptions
• Model estimation and fit
• Interpreting results
• Validation

5
2.1. Design

• Method
• Select attributes (number, type)
• Choose model form (additive? dependent variable?)
• Individual or aggregate estimation?
• Traditional, Choice-based or Adaptive conjoint?

6
2.1. Design

• Stimuli Factor ( Attribute) selection
• Criteria
• Differentiate
• Able to communicate
• Actionable
• Price ? Could enter as separate attribute, mind
  correlations or infeasible stimuli
• Levels
• Strive for Balance
• Range Feasible, Relevant, Stretch

7
2.1. Design

• Stimuli Utility specification
• Part worth, Ideal Point or Linear model?
• Main effects or interactions?

8
Alternative Models
9
2.1. Design

• Data collection
• Presentation
• Trade-off
• Full profile (Fractional factorial)?
• Preference Measure
• Ranking
• Rating
• Choice (no-)
• Task per respondent (Regular, Adaptive, Hybrid?)

10
Example Sneakers

• 3 attributes, 3 levels each
• Sole Rubber, Polyurethane, Plastic
• Upper Leather, Canvas, Nylon
• Price 30, 60, 90
• Fractional Factorial 9 out of 27 profiles (3
  sole x 3 upper x 3 price) evaluated

3
2
1
1
2
3
1
2
3
11
Example Profiles for Sneakers
Stimulus Sole attribute 1 Upper attribute 2 Price attribute 3
1 Rubber (1) Leather (1) 30 (1)
2 Rubber (1) Canvas (2) 60 (2)
3 Rubber (1) Nylon (3) 90 (3)
4 Polyurethane (2) Leather (1) 60 (2)
5 Polyurethane (2) Canvas (2) 90 (3)
6 Polyurethane (2) Nylon (3) 30 (1)
7 Plastic (3) Leather (1) 90 (3)
8 Plastic (3) Canvas (2) 30 (1)
9 Plastic (3) Nylon (3) 60 (2)
(attribute level)
12
2.2. Assumptions

• Few statistical assumptions
• Theory-driven design, estimation and
  interpretation
• Overfitting?
• GIGO (Garbage in Garbage out)?

13
2.3. Model Estimation and Fit

• E.g. Additive Model, part-worths
• where U(X)utility of alternative X, m
  attributes, kiattribute levels of attribute i,
  xij1 for level j of i, 0 elsewhere, ?ijpart
  worth for level j of i
• Bv (Usneakers2) ?11 ?22 ?32

14
2.3. Model Estimation and Fit (ctd)

• Purpose Find levels of ?ij that reflect
  consumers stimuli evaluations as closely as
  possible
• Method
• Ranking MONANOVA, Linmap
• Rating Dummy-variable regression
• Choice MNL or Probit model
• Fit
• Correlate actual/predicted ranks
• Hit rate
• R2

15
Example Profiles for Sneakers
Stimulus Sole attribute 1 Upper attribute 2 Price attribute 3
1 Rubber (1) Leather (1) 30 (1)
2 Rubber (1) Canvas (2) 60 (2)
3 Rubber (1) Nylon (3) 90 (3)
4 Polyrethane (2) Leather (1) 60 (2)
5 Polyrethane (2) Canvas (2) 90 (3)
6 Polyrethane (2) Nylon (3) 30 (1)
7 Plastic (3) Leather (1) 90 (3)
8 Plastic (3) Canvas (2) 30 (1)
9 Plastic (3) Nylon (3) 60 (2)
(attribute level)
16
2.3. Model Estimation and Fit (ctd)

• Example Sneakers Preference ratings and Variable
  Indicator Coding (last level Base)

Preference Rating
Rubber
Poly
Leather
Canvas
60
Sneaker
30
1 2 3 4 5 6 7 8 9
Sole
Upper
Price
17
2.3. Model Estimation and Fit (ctd)
18
2.4. Interpreting results

• Assess part-worths for attribute levels
• Evaluate attribute importance
• Use choice simulator

19
Assess part-worths for attribute levels

• Example Indicator Coding, AttributeSole
• b11 coëfficiënt Sole11
• b12 coëfficiënt Sole2-.333
• b130
• Average (1-.3330)/3.222
• Calculate part worths such that sum 0?
• -gt ?11 b11-Average1-.222. 778
• ?12 b12-Average-.333-.222?-.556
• ?13 b13-Average-.222

20
Example Sneakers Outcome Part worth calculations

• Sole ?11.778, ?12 -.556, ?13 -.222
• Upper ?21.445, ?22 .111, ?23 -.556
• Price ?311.111, ?32 .111, ?33-1.222

21
Part Worths Sneakers
22
Evaluate attribute importance
where iattribute, j attribute level, m number
of attributes, Ii range of part worths for
attribute, Wi attribute importance (share)
23
Attribute importance

• Example Sneakers
• Sole .286
• Upper .214
• Price .5

100
60
24
Calculating Attribute importance
25
2.5. Validation

• On holdout sample?
• Clusters of respondents
• Alternative Models?
• Significance (overfitting)?

26
3. Case

• Channel and Price Offers for Safety Products

27
Problem Statement

• A company specialized in safety-related products,
  intends to improve its channel- and pricing
  approach for different types of products.
• Preferred combination, by consumers, of
  information channel, selling channel, and price
  level?

28
Problem Statement (ctd)

• Consumers can obtain information, and/or purchase
  products,
• through the internet (companys website)
• from a safety consultant /advisor (in home)
• in BM stores
• Prices can deviate from a recommended price

29
Research Setup

• Use conjoint analysis to assess consumer
  preference for alternative channel/price
  combinations
• Conduct analysis for three types of products
• Bicycle Lock
• Fire Blanket
• Alarm system

30
Design Stimuli

• Attributes
• Utility Part worths, additive

31
Design Data Collection

• Traditional Method
• Full Profile approach
• 27 possible combinations fractional, orthogonal
  design -gt 9 profiles/product/respondent
• Preference measure rating
• Respondent task regular, 2 products

32
Data Collection (ctd)

• Info products/recommended prices
• (e.g. fire blanket 46.05Euros, Alarm system
  315.70Euros, )
• Info channels
• BM store (where, what, chain)
• Internet (site, what)
• Advisors where, education/expertise

33
Scenario (Stimulus) 1

• Imagine
• You use the internet to gather information on the
  fire blanket
• You purchase the fire blanket in the store
• The recommended price is 46.05Euros
• In the store, you pay this recommended price 10
• How do you rate this scenario? ./100

34
Model and Variable Coding

• Dataset see File Caseconj.sav
• Cases respondentsprofiles
• Dummy variable regression per product and across
  respondents,
• dependent rating
• Independent 6 dummy variables (TI, TA, II, IA,
  PR, PL) reference scenario transaction and info
  in BM, higher price.

35
Estimation Results

• See output file Caseconj.spo

36
Interpretation

• Part Worths and Attribute importance
• E.g. Fire Blanket
• Information channel no significant impact
• Transaction channel (.365)
• Internet -7.78, Advisor -.1, Store 7.88
• Price (.635)
• Low 15.22, Medium 2.83, High 12.38

37
Validation

• Estimation Sample Correlation between true and
  predicted scores? (Fire Blankets .435)
• Holdout sample
• Re-estimate and compare coefficients?
• Correlate true and predicted scores in holdout

38
Outcome

• Attribute importance?
• E.g. Bicycle Lock First price (27.6), then
  transaction channel (15.7), info channel not
  important (1.5)
• Most appealing offer customer
• E.g. Bicycle Lock Store, Low price. Utility
  7.88 15.23 23.11
• Trade off e.g. Bicycle Lock
• Store, medium price 7.88-2.835.05
• Internet, low price -7.7815.227.44
• Prefer latter option!

39
Outcome (ctd)

• Customer heterogeneity?
• E.g. Male vs female
• Individual analysis?
• Product differences in attribute significance,
  importance, part worths!
• E.g. Best info channel depends on product
  Bicycle Lock store, Alarm system advisor